Search

• Proposed FSDnet has good performance in crop plant phenotype segmentation. • Introduced Gaussian density can reduce the calculation and improve accuracy. • By embedding FSDnet, the traditional models can work well in different densities. • We construction a new field 3D deep learning dataset including strawberry and apple. The accurate segmentation of fruit phenotypes in the field is of great significance for agricultural automation in the 3D scene. Although the existing fruit segmentation based on 3D point cloud has made great progress, in the complex field environment, due to lighting, leaf occlusion, shooting angle and other problems, the point cloud obtained by depth camera often has the problem of multiple voids and discrete points, which seriously affects the accurate segmentation of fruit phenotype. This paper proposes a embedding subnetwork FSDnet based on density-based feature extraction and feature propagation and embeds it in the novel segmentation networks, which effectively improves the segmentation accuracy of the point cloud phenotype in multi-hole and multi-discrete fruits, including (1) The density-based point cloud feature extraction and feature propagation theory is proposed to alleviate the problem of perception degradation in fruit edge point caused by discrete points and holes caused by imcomplete point cloud in the agriculture scene. (2) A density-adaptive embedding semantic segmentation framework FSDnet is proposed, and embedding the classical point cloud neural network can significantly improve the segmentation accuracy of the fruit phenotypes with multiple holes and discrete points in the traditional network. (3) This paper made a strawberry dataset and tested the designed new neural network on both strawberry and apple filed dataset. After FSDnet is embedded on different novel net, almost all net have been improved. We verified the performance of FSDnet in different density states in agricultural scenarios, mitigated the negative impact of density on segmentation accuracy, proving that it can adapt to different point cloud density in agricultural scenarios in comparison between Gaussian density and other two traditional density schemes, Gaussian density reduces the computational traffic (0.58G) of the network while maintaining similar performance to the other two densities, proving the superiority of assuming a Gaussian density. [ABSTRACT FROM AUTHOR]

• A bibliometric and content analysis was conducted to comprehensively review and analyze deep learning research in cropland field identification. • The paper discusses the challenges of deep learning-based research on cropland field identification, providing readers with insights and directions for future research in the field • This study fills the gap by providing a systematically summarized review of this research area. The cropland field (CF) is the basic unit of agricultural production and a key element of precision agriculture. High-precision delineations of CF boundaries provide a reliable data foundation for field labor and mechanized operations. In recent years, with the dual advancements in remote sensing satellite technology and artificial intelligence, enabling the extraction of CF information on a wide scale and with high precision, research on CF identification based on deep learning (DL) has emerged as a highly esteemed direction in this field. To comprehend the developmental trends within this field, this study employs bibliometric and content analysis methods to comprehensively review and analyze DL research in the field of CF identification from various perspectives. Initially, 93 relevant literature pieces were retrieved and screened from two databases, the Web of Science Core Collection and the Chinese Science Citation Database, for review. The previous studies underwent quantitative analysis using bibliometric software across five dimensions: publication year, literature type and publication journal, country, author, and keyword. Subsequently, we analyze the current status and trends of employing DL in the field of CF identification from four perspectives: remote sensing data sources, DL models, types of CF extraction results, and sample datasets. Simultaneously, we combed through current publicly available sample datasets and data products that can be referenced to produce sample datasets for CFs. Finally, the challenges and future research focus of DL-based CF identification research are discussed. This paper provides both qualitative and quantitative analyses of research on DL-based CF identification, elucidating the current status, development trends, challenges, and future research focuses. [ABSTRACT FROM AUTHOR]

In current plant phenotyping research, the study of plant time-series images based on deep learning has received widespread attention. While such image data is relatively easy to obtain, the cost of annotation is high. One efficient method for achieving cost-effective training is through contrastive learning. Plant growth is slow, and the changes in image sequences over a period of time are small, with simple semantic information. Previous contrastive pre-training models struggled to effectively distinguish positive samples from the same image with different augmented views and similar negative samples from different images. Therefore, this paper proposes a method called self-supervised contrastive learning method for plant time-series images with a Priori Distance Embedding (PDE). The semantic information in images corresponding to different phenological stages of plants varies. This method transforms this crucial domain knowledge into prior distances for image pairs and conducts contrastive learning pre-training. The learned weights can be transferred to downstream tasks. Building upon this method, experiments were conducted on cherry time-series images to assess the quality of pre-training through a plant phenotyping image semantic segmentation task. To provide a comprehensive example of plant time-series image phenotypic analysis, this paper establishes a cherry growth temporal model, specifically including PDE pre-training, anomaly detection, semantic segmentation, and recording the results from the temporal dimension. The experiments indicate that this self-supervised contrastive learning method can be effectively applied to the pre-training of plant time-series images, demonstrating broad applicability in various computer vision studies related to plant phenotyping. • A pre-training method for plant images with a Priori Distance Embedding (PDE). • Transforms the phenological information into prior distances for image pairs. • An image acquisition device was built to obtain cherry time series images. • The PDE method was tested by plant phenotype image segmentation task. • We provided an exhaustive example of cherry time-series image phenotypic analysis. [ABSTRACT FROM AUTHOR]

• An advanced measurement method enables precision farming. • A precise, stable and novel live chickens shank length measurement method. • Fused images provide richer features. • An improved ResNet model applied to a regression task. Phenotypic parameters are crucial reference indicators in poultry breeding. However, the chickens shank length is still manually measured, which is time-consuming and labor-intensive. Additionally, the measurement results are difficult to unify due to the subjective factors of different individuals. To address this issue, this paper proposed a method for live chicken shank length measurement (SLM). It enriches chicken shank feature by fusing visible images and infrared images. The fusion images are then input into a deep regression model based on the improved ResNet. The measurement model used ResNet as its backbone and introduces Squeeze-and-Excitation (SE) blocks and a Spatial Pyramid Pooling (SPP) block, resulting in more precise and stable shank length measurements. The average coefficient of variation, average floating error, average standard deviation and Pearson correlation coefficient for shank length measurements using the fusion images are 0.21 %, 0.49 %, 0.181 mm and 0.996, respectively, compared with using single visible or infrared image, the accuracy and stability are obviously improved. That indicated combining deep learning model and fusion information, the SLM proposed in this paper can achieve a more precise, reliable and standardized measurement of live chicken shank length. [ABSTRACT FROM AUTHOR]

• We proposed a delineating field parcel model (TSANet) based on satellite image time series. • TSANet learn the relevance of spatial-spectral-temporal feature representation. • TSANet is robust across space and time. • TSANet performs better than pervious methods. • The time series data from the main growing period has a greater impact. Satellite image time series (SITS), such as Sentinel-2 imagery, plays a crucial role in the delineation of agricultural fields by reducing the impacts of ambiguities due to the spatial arrangement of field boundaries. Existing delineate field parcel models rely extensively on spatial features derived from single-date imagery. However, several studies have exploited the potential of SITS to effectively tackle the complexities associated with the intrinsic consistency between agricultural fields and their boundaries. This paper proposes a novel Two-Stream Attention convolutional neural network (TSANet) to capture the subtle difference between agricultural fields from SITS. Specifically, a field temporal semantic stream is introduced to adaptively leverage the significance of spatial-spectral-temporal feature representation associated with the location of agricultural parcels, especially where transitions in crop types take place. Considering the consistency between field parcels and their boundaries, we developed a field boundary prediction stream to enhance the extraction of edge features, particularly for the extraction of small and irregular agricultural parcels. Moreover, a field parcel refining block is employed to further enhance the geometric accuracy of agricultural fields. We conducted experiments on Sentinel-2 images from the Netherlands. Results showed that our approach produced a better layout of agricultural fields, with an average F1-score of 0.91 than the existing 3D-UNet, U-TEA, and BiConvLSTM. In addition, through the analysis of both quantitative and qualitative results, the stronger robustness of the model compared to other algorithms has been verified by temporal transfer and large-scale spatial prediction. We compared the difference between SITS and the corresponding composite images, which further verified the influence of temporal variation on the proposed approach. This paper provides a general guide for delineating agricultural parcels using SITS. [ABSTRACT FROM AUTHOR]

Soybean is a critical source of food, protein, and oil, and thus has received extensive research aimed at enhancing their yield, refining cultivation practices, and advancing soybean breeding techniques. Within this context, soybean pod counting plays an essential role in understanding and optimizing production. Despite recent advancements, the development of a robust pod-counting algorithm capable of performing effectively in real-field conditions remains a significant challenge. This paper presents a pioneering work of accurate soybean pod counting utilizing unmanned aerial vehicle (UAV) images captured from actual soybean fields in Michigan, USA. Specifically, this paper presents SoybeanNet, a novel point-based counting network that harnesses powerful transformer backbones for simultaneous soybean pod counting and localization with high accuracy. In addition, a new dataset of UAV-acquired images for soybean pod counting was created and open-sourced, consisting of 113 drone images with more than 260k manually annotated soybean pods. The images are taken from an altitude of approximately 13 ft, with angles between 53 and 58 degrees, under natural lighting conditions. Through comprehensive evaluations, SoybeanNet demonstrates superior performance over five state-of-the-art approaches when tested on the collected images. Remarkably, SoybeanNet achieves a counting accuracy of 84.51% when tested on the testing dataset, attesting to its efficacy in real-world scenarios. The publication also provides both the source code and the labeled soybean dataset, offering a valuable resource for future research endeavors in soybean pod counting and related fields. • Created and open-sourced the 1st UAV-acquired image dataset for soybean pod counting. • Pioneered the first study on precise soybean pod counting utilizing UAV images. • Demonstrated SoybeanNet's superiority over 5 state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

• An efficient method is proposed for orchard tree trunk detection in natural scenes. • A multi-scenes orchard tree trunk dataset is established for experiments. • The YOLOv5 model structure is improved to enhance detection accuracy and speed. • Proposed method is robust against complex orchard environment. The recognition and localization of fruit tree trunks in orchard are important for orchard operation robots, which are the bases for automatic navigation, fruit tree spraying and fertilization etc. A method was proposed based on machine vision to detect target objects such as fruit tree trunks, person and supporters in orchard by improving the YOLOv5 deep learning algorithm in this paper, which is applicable to the recognition tasks in natural orchard environments. Firstly, 1354 images of the natural orchard collected by camera were image enhanced, weather effects such as rain, snow, bright light, shadow and fog were added to expand the dataset and to increase the robustness of the model. Secondly, the original YOLOv5 model was improved by replacing the Bottleneck network in the C3 module with the lightweight GhostNet V2 to reduce the network parameters, and changing the box loss function CIoU to SIoU in the loss function to make the regression of the detection box more accurate, and coordinate attention mechanism (CA) was added to the network to reduce the interference of useless background information in images. Before training, pre-anchor boxes were generated by using IoU-based K-means clustering, after that the dataset was fed into the improved YOLOv5 for training, and the trained model was used to detect the trunks. Finally, weighted boxes fusion (WBF) was used instead of the non-maximum suppression (NMS) in this paper for the output of the detection boxes. Then the density-based spatial clustering of applications with noise (DBSCAN) algorithm was used for trunk clustering. The improved target detection method was trained and validated on the experimental dataset. The model size is reduced by 43.6 %, model parameters are reduced by 46.9 %, and the mAP reaches 97.1 %, with an average detection speed of 198.2 ms per image. Compared with the original YOLOv5, the model is more lightweight, the detection accuracy and speed are improved. The improved YOLOv5 is also better than YOLOv3, NanoDet and SSD in terms of combined accuracy and speed, and has similar performance to YOLO_MobileNet in orchard dataset. The experimental results show that the improved YOLOv5 target detection model proposed in this paper is lightweight while still having better detection accuracy and detection speed in complex environments, and the model is small enough to be deployed to mobile or low-performance terminals for target detection in natural orchard environments. [ABSTRACT FROM AUTHOR]

With the development of modern animal husbandry, especially pig farming, large-scale, intensive, and automated farming has become the trend in the industry. The realization of accurate recognition and warning of dangerous actions in feeding scenarios for sows and piglets holds high research and practical value in this field. This paper addresses this issue by proposing a Transformer-based Neural Network (TNN) model. This model emphasizes the extraction of global features and handling of long-distance dependencies, significantly improving the accuracy of behavior recognition. Compared with traditional neural network models, TNN demonstrates superiority in dealing with animal behavior recognition in farming scenarios. Furthermore, an in-depth study of the attention mechanism in the TNN model was conducted in this paper. By visualizing the attention heat map of the TNN model, it was found that TNN could effectively focus on key areas in the image, thereby accurately identifying the behavior of piglets. Finally, this paper proposes a unique model lightweighting strategy that allows the TNN model to run efficiently on edge devices. In the experimental part, the performance of the TNN model on five behavior recognition tasks was evaluated first. The results showed that the TNN model achieved high scores in both Precision and Recall, far exceeding traditional neural network models. Then, by visualizing the attention heat map of the TNN model, the superiority of the TNN model in focusing on key image areas was further confirmed. In the end, the effect of the model lightweighting strategy was demonstrated, and even with a significant reduction in parameters and computational complexity, the performance of the TNN model remained excellent. This research not only promotes the development of animal behavior recognition technology but also provides new insights and tools for the precise management and efficient operation of the animal husbandry industry. • A comprehensive dataset of sow nursing scenarios was collected and annotated. • A variety of data augmentation methods were proposed to rebalance it, ultimately leading to a model accuracy of 0.91 mAP. • A time-series neural network (TNN) model is proposed, tailored to the characteristics of the actual farm collection equipment. • An alarm triggering mechanism was designed, enabling the model's application in actual agricultural scenarios. [ABSTRACT FROM AUTHOR]

• Calyx localization method proposed in this paper obtained the best localization. • YOLOv5x improved detection accuracy of feature points in binocular images. • Pairing of kiwifruit and calyx facilitated calyx matching for kiwifruit localization. • Calyx matching mechanism effectively avoided feature points mismatching in global matching method. • Template matching with parallel polar line constraint improved success rate of kiwifruit matching. Localization is the first critical step for picking robots to successfully grasp fruit. However, classical binocular localization methods adopted global matching for kiwifruit, which may result in a large amount of mismatching feature points in complex orchard and thus cause low localization accuracy. Therefore, an improved binocular localization method of calyxes based on deep learning was proposed to accurately detect and locate kiwifruit for robotic harvesting. Calyxes in the binocular images and kiwifruit in the left images of the binocular images were detected using You Only Look Once version 5 xlarge (YOLOv5x). The detected calyxes were matched in the binocular images using kiwifruit and calyx pairing and kiwifruit matching. The matched calyxes in the binocular images were used to locate calyxes using three localization methods. Specifically, three binocular localization methods, i.e., calyx localization (CL), fruit localization (FL), and depth information from depth map (DIDM), were compared to find the optimal one. Ground truth three-dimensional coordinates of calyxes was measured by laser rangefinder and coordinate paper on a self-designed experimental platform. Results showed that YOLOv5x achieved an average precision (AP) of 99.5 % on kiwifruit detection and a mean AP of 93.5 % on kiwifruit and calyx detection with a detection speed of 108 ms per image. Average deviation of X-axis, Y-axis, and Z-axis obtained by the CL method were 7.9 mm, 6.4 mm, and 4.8 mm, respectively. Compared with the FL and DIDM methods, localization error rate of the proposed CL method was reduced by 55.1 % and 53.8 %, respectively. It indicates that the proposed CL method is promising for robotic harvesting. [ABSTRACT FROM AUTHOR]

• This paper establishes an image dataset of sika deer obtained from different perspectives. • This paper proposes an AD-YOLOv5 model to detect key parts of the sika deer. • This paper improves the loss function of the model to enhance both the training and inference ability of the model. • This paper proposes an adaptive channel attention mechanism for better extracting shallow features of the key parts of sika deer. • By altering the feature fusion network and combining it with the channel attention mechanism, the feature information is balanced across different scales, which enhances the accuracy of detection. Accurate and rapid fixation of key parts of sika deer is essential for the mechanized and intelligent collection of both deer blood and velvet antler. To solve the low detection accuracy problem of key parts of deer caused by the diversity of deer farm breeding environment and the rapid movement of deer, this paper proposes an AD-YOLOv5 algorithm for detecting key parts of sika deer based on YOLOv5s. Specifically, a new weighted bidirectional feature pyramid network S-BiFPN is first proposed, which both simplifies the structure and effectively performs skip feature fusion to better capture object information at different scales, so that the adaptive feature fusion capability of the model is optimized. Second, this paper proposes to add a SENet module after the C3 module of the 9th Layer in the backbone part of the network, so that the network can adaptively readjust the channel weights of feature maps to enhance the important features for key parts of sika deer and improve the expression ability of the model for key parts. Furthermore, to make the model more stable in the training process, a novel bounding box regression loss function SIoU is led into, which can better learn the position and size of the bounding box by fusing the orientation information between the ground truth box and the predicted box. The effectiveness of the proposed algorithm is demonstrated through ablation experiments. Experimental results show that the mAP @0.5 of the proposed algorithm reaches 97.30 % for the detection of key parts of sika deer, which is 12.85 %, 9.65 %, 9.18 %, 7.13 %, 3.59 %, 4.6 %, 0.90 % and 0.10 % higher than SSD, EfficientDet, YOLOv3, YOLOv4, Faster R-CNN, YOLOv5s, YOLOv7 and YOLOv8s, respectively. This research provides a new idea for detecting key parts of animals in complex environments and lays a theoretical foundation for the development of intelligent fixation devices for sika deer. [ABSTRACT FROM AUTHOR]

[Display omitted] • An image-based system for monitoring pests in mango orchards was developed. • System overcame bottlenecks of pest monitoring in remote and outdoor environment. • Deep learning algorithm detects and recognizes mango pests with an F 1 -score of 0.96. • More than 2 years of spatiotemporal data reflected trends in pest behavior. • Spatiotemporal data was used to realize new data-driven IPM strategies. Mango production is a prominent tropical fruit industry worldwide. However, outdoor mango cultivation is susceptible to crop damage caused by insect pests and harsh environmental conditions. Integrated pest management (IPM) has emerged as a proposed solution to this problem. IPM utilizes data-driven and environmentally-friendly methods to suppress insect pest populations. Nevertheless, the collection of insect pest population data remains a laborious process, necessitating automation. This paper presents an image-based monitoring system to automatically record insect pest populations and environmental conditions in mango orchards. The system comprises solar-powered sensor nodes capable of periodically acquiring and analyzing sticky paper trap images. A modular deep learning-based algorithm was developed to detect and classify insect pests into seven classes, including major insect pests of mango such as thrips, mango leafhopper, and oriental fruit fly, with an average classification F 1 -score of 0.96. Unlike other insect counting algorithms, the algorithm reliably classifies insect pests according to different taxonomic levels even in non-laboratory environments. The monitoring system was tested and deployed in a remote mango orchard for over two years. The collected spatiotemporal information was analyzed to demonstrate the benefits of using the proposed system and recommend new IPM strategies. Temporal data analysis revealed a significant decrease in the count of selected insect pests after using the system, enabling identification of insect hotspots through statistical methods. This work presents a breakthrough in hardware and software solutions for developing smarter insect pest monitoring systems, leading to better IPM strategies. [ABSTRACT FROM AUTHOR]

• Precision agriculture method applicable to remote on-tree pomegranate pixel-based fruit size monitoring. • Comparison of Deep Learning and zero-shot vision Transformer instance segmentation models. • Image-Based size estimation algorithm for occluded fruits with protuberances. In situ automatic fruit monitoring is of great interest for more accurate and cost-efficient decision making in agriculture. For this purpose, the development of computer vision-based tools is essential. Deep Learning techniques have shown good performance in fruit detection and segmentation. Recently, new models based on Transformer architecture have emerged with promising potential and zero-shot inference capability. In this paper, a Deep Learning model, Mask R-CNN, was trained for on-tree pomegranate fruit segmentation and compared with foundational models based on Vision Transformer, Grounding DINO and Segment Anything Model. Results with Mask R-CNN proved a better performance, according to F1 score and AP metrics, and a lower computational cost, according to prediction time. One of the most interesting derived applications from fruit segmentation is fruit size estimation. However, segmented fruit masks are frequently incomplete due to occlusions. Therefore, image fruit size estimation is not a straightforward process. In this work, we also propose a novel algorithm to estimate and monitor the fruit size in pixel units from the automated masks. A median relative error of 1.39% was obtained, demonstrating the potential and feasibility of future fully-automatic fruit size estimators. [ABSTRACT FROM AUTHOR]

• A method for estimating leaf density of fruit-tree canopies based on the wind-excited audio. • Developed a spectral centroid attention based on the spectral-centroid distribution probability of wind-excited audio. • DCNN-based accurate estimation of leaf-densities of fruit-tree canopies. It is essential for precision air-assisted ground sprayer to accurately detect the leaf density of fruit trees in agriculture. However, it is difficult to achieve real-time measurement of canopy leaf densities by using conventional sensors (such as LIDAR and camera). In fact, the interaction between wind and canopies of different leaf densities may generate different excited-audios during air-assisted spraying. Therefore, this paper proposes a method for leaf-density detection of fruit-tree canopies based on the variance of wind-excited audios. First, the fused spectrogram (FSP) was constructed by the short-time fast Fourier transform (STFT), consisting of spectral features and the spectrogram of audio signals. Then, the estimation model between the FSP and leaf densities was developed by using deep convolutional neural network (DCNN), which was enhanced by a spectral centroid attention (SCA) based on the distribution function of spectral centroids. Finally, the test results showed that: (1) the developed model achieved a 96.93% accuracy and a 97.96% precision in leaf-density recognition, and (2) compared with unimproved DCNN model, the accuracy and precision of the developed model were increased about 7.85% and 8.47%, respectively, which indicated that this method could achieve the prediction of leaf-density based on wind-excited audios. The study is expected to provide a reference for leaf-density detection of fruit tree canopies and real-time control of precision spray. [ABSTRACT FROM AUTHOR]

• Novel U-LSTM and FADA-A modules are introduced into FieldSeg-DA framework. • U-LSTM merges spatiotemporal features in multi-resolution satellite imagery. • FADA-A aligns cross-domain features with guidance of land cover product. • U-LSTM and FADA-A enhance temporal and spatial transferability respectively. Deep learning has become the leading technique for precisely extracting individual arable fields (IAFs) from high-resolution remote sensing images. Maintaining transferability remains a major concern for deep learning methodologies due to the significant cost of acquiring labelled samples for network training. FieldSeg-DA, introduced by Liu et al. (2022) , enhanced network transferability by introducing a finely tuned adversarial domain adaptation module (FADA). An improved version of the FieldSeg-DA framework, FieldSeg-DA2.0, which further enhances transferability through incorporating multisource remote sensing and land cover data is introduced in this paper. First, we introduce a spatiotemporal fusion module, U-LSTM, to extract the IAF extent by merging textural information from the high-resolution image (Gaofen-2) and phenological information from the coarse time-series data (Sentinel-2). Incorporating phenological information mitigates the risk of overfitting associated with the high-resolution imagery acquired in specific seasons, thereby improving the temporal transferability of FieldSeg-DA2.0 compared to FieldSeg-DA. Second, we introduce a novel fine-grained adversarial domain adaptation module with ancillary data (FADA-A) to enhance spatial transferability. FADA-A incorporates prior knowledge from the Dynamic World (DW) land cover dataset to guide adversarial training in the standard FADA, thereby enhancing the robustness of domain adaptation across diverse geographic regions. We evaluate the performance of FieldSeg-DA2.0 through various spatial and temporal transfer experiments utilizing GaoFen-2 and Sentinel-2 data. The results illustrate that the cross-domain performance of FieldSeg-DA2.0 is significantly better than that of the original FieldSeg-DA, highlighting its robust spatiotemporal transferability. FieldSeg-DA2.0 can accurately delineate IAFs across varied regions and seasons without requiring additional training samples, illustrating its considerable potential for large-scale IAF extraction. [ABSTRACT FROM AUTHOR]

The past decade has witnessed the rapid development and adoption of machine and deep learning (ML & DL) methodologies in agricultural systems, showcased by great successes in applications such as smart crop management, smart plant breeding, smart livestock farming, precision aquaculture farming, and agricultural robotics. However, these conventional ML/DL models have certain limitations: they heavily rely on large, costly-to-acquire labeled datasets for training, require specialized expertise for development and maintenance, and are mostly tailored for specific tasks, thus lacking generalizability. Recently, large pre-trained models, also known as foundation models (FMs), have demonstrated remarkable successes in language, vision, and decision-making tasks across various domains. These models are trained on a vast amount of data from multiple domains and modalities. Once trained, they can accomplish versatile tasks with just minor fine-tuning and minimal task-specific labeled data. Despite their proven effectiveness and huge potential, there has been little exploration of applying FMs to agriculture artificial intelligence (AI). Therefore, this study aims to explore the potential of FMs in the field of smart agriculture. In particular, conceptual tools and technical background are presented to facilitate the understanding of the problem space and uncover new research directions in this field. To this end, recent FMs in the general computer science (CS) domain are reviewed, and the models are categorized into four categories: language FMs, vision FMs, multimodal FMs, and reinforcement learning FMs. Subsequently, the process of developing agriculture FMs (AFMs) is outlined and their potential applications in smart agriculture are discussed. In addition, the unique challenges and risks associated with developing AFMs are discussed, including model training, validation, and deployment. Through this study, the advancement of AI in agriculture is explored by introducing AFMs as a promising paradigm that can significantly mitigate the reliance on extensive labeled datasets and enhance the efficiency, effectiveness, and generalization of agricultural AI systems. To facilitate further research, a well-classified and actively updated list of papers on AFMs is organized and accessible at https://github.com/JiajiaLi04/Agriculture-Foundation-Models. • Basics of large language and foundation models. • Review of potential applications of large language and foundation models in agriculture. • Outline challenges and opportunities. [ABSTRACT FROM AUTHOR]

• The model can detect abnormal manure and locate the cage of sick hens. • The YOLOv5 model was used to detect abnormal manure. • SuperPoint-SuperGlue was used for distance estimation. • A method for reducing cumulative errors in distance estimation. • The backtracking accuracy of the model reaches 91.6%. Digestive diseases are common poultry diseases that significantly affect production and animal welfare. Early warning of digestive disease outbreaks is essential in poultry breeding. In conventional cage laying hen systems, breeders usually spend 1–2 days randomly selecting hens for health assessment. This process is time-consuming, labor-intensive, and may cause stress to the hens, which goes against animal welfare principles. In this paper, a model structure was proposed to locate the cage of sick hens based on manure images using digital image processing and deep learning. Abnormal features of poultry manure was extracted using the YOLOv5 and combined with the motion information of the manure belt estimated by the SuperPoint-SuperGlue model to construct an early warning and traceability system for diseases in a layered cage farming mode. This system enables the automatic detection and traceability of digestive diseases in layer-raised hens. The results showed that YOLOv5 can effectively recognize abnormal chicken manure with precision rate, recall rate, mAP@0.5 and mAP@[0.5:0.95] of 95.7 %, 95.4 %, 98.1 % and 84.3 %. The SuperPoint-SuperGlue model demonstrated strong data fitting capabilities and remarkable robustness in long-distance monitoring of manure belt movement, with MAE, RMSE, and MAPE values of 0.19 m, 0.14 m, and 0.34 %, respectively. After combining the SuperPoint-SuperGlue with a self-designed error correction module, the traceability accuracy of the model was validated through 48 field tests, achieving a correct tracing rate of 91.6 %. These research results can be used for providing early warnings of digestive system diseases in caged chickens, providing a prerequisite for the development of intelligent disease diagnostic equipment. [ABSTRACT FROM AUTHOR]

AgroCounters is an open-source repository for counting objects in images in the agricultural domain by utilizing deep-learning algorithms. In this paper, we present the framework of AgroCounters, which integrates state-of-the-art deep learning models, including regression-based counting, detection-based counting, and density-estimation-based counting, to accurately count various agricultural objects, such as fruits, vegetables, and livestock, in single images. The framework utilizes transfer learning techniques to optimize model performance on the limited labeled data available in the agricultural domain. We provide an open-source implementation of AgroCounters, which includes a multitude of algorithms for counting applications and a toolbox that includes metrics, training data tools, visualizations, and a simple installation guide for several open-source implementations of counting methods. We evaluated the performance of AgroCounters on multiple agricultural datasets acquired from RGB sensors, including plant leaves, melons, wheat grains, cherry tomatoes, grapes, apple flowers, bananas (fruit and leaves), pears, and chickens. We compared the results of the various implemented methods over these datasets and showcased the most suitable solution for each. YOLOv5, the most recent of the compared object detectors, provided the best results on all the examined datasets, and there was no clear 'winner' between Faster-RCNN and RetinaNet. Based on the analyzed datasets, when higher accuracy is required, the direct regression network (DRN) should be used; for small datasets, multiple scale regression (MSR) gives superior results. Based on the developments, we proposed guidelines for developing deep-learning-based counting solutions for agricultural applications, focusing on solutions and best practices for the agricultural domain. Overall, AgroCounters presents a promising solution for automated counting in the agricultural domain, offering significant potential for reducing manual labor, improving crop management, and increasing productivity. • AgroCounters is a repository for counting agriculture objects using deep learning. • AgroCounters includes regression, detection, and density estimation algorithms. • AgroCounters provides toolboxes for metric calculation, training, and visualization. • AgroCounters publicizes 10 different datasets with annotations and trained models. • Present guidelines for developing deep learning-based counters in agriculture. [ABSTRACT FROM AUTHOR]

• An interpretable machine learning method is developed for the SPAC system. • A local Gaussian process group is built to simulate multiple SPAC subprocesses. • Improved kernel function can interpret the key mechanisms learned by machine. • Improved kernel function can capture nonstationary dynamics during crop growth. • Our method leads to advanced prediction accuracy and generalization ability. Traditional models of the soil-plant-atmosphere continuum (SPAC) system are physics-based and their practical application has been hindered by issues of high parameterization, prior structural bias, and costly running. In this paper, we developed a machine learning modelling method based on the Gaussian process (GP) to avoid these difficulties of traditional modelling. At the same time, shortcomings of conventional machine learning in terms of interpretability and mechanism representation were addressed. The innovative method consisted of structuring a local GP group framework and improving kernel functions used in each local GP. The resultant model was endowed with advanced capacities, including representing interplays between subprocesses within the complex SPAC system, representing nonstationary subprocess dynamics caused by crop growth stage shifts, as well as automatically interpreting key low-order input interactions and dominant input variables for each subprocess. The performance of our model was examined on synthetic SPAC system datasets that covered three different soil conditions. Results demonstrated that its interpretations regarding subprocess mechanisms were robust across different soil conditions and consistent with domain knowledge. Compared to a conventional global GP model and two deep learning models (DNN and LSTM), our model performed significantly better not only in regular prediction experiments but also in generalization experiments that required models to be transferred across different conditions. Our study suggested that it is promising to enable a machine learning model interpretable by improving its feature representation function. For machine learning modelling of complex systems like the SPAC, it may be more important to enhance the model's mechanism representation capability than simply using deeper networks. [ABSTRACT FROM AUTHOR]

Smart monitoring systems in orchards can automate agriculture monitoring processes and provide useful information about the presence of insects, such as the Brown Marmorated Stink Bug (BMSB), that threaten the production quantity and quality of fruit such as pears. Unlike other approaches in the literature, we propose a low-cost image monitoring system which exhibits a very low power consumption without compromising much of the accuracy that existing expensive systems incorporating significant computing and processing capability can achieve in such applications. The proposed system relies on a microcontroller unit and a camera which can take pictures of a double-sided sticky insect trap which, with the help of novel machine learning algorithms, can report on the presence of BMSB via a long-range communication link. The Internet of Things data capture and analysis system has recently been deployed in a real orchard in Italy which is subject to BMSB infestation and the first images have been analysed. This paper presents how the system works, the image processing, detection and classification algorithms, as well as a demonstration of the memory and energy consumption associated with the processing algorithms. The system achieves an accuracy of over 90% with multiple times less memory and energy consumption compared to other similar approaches in the literature. • Low-Power Edge-based insect monitoring system. • Insect detection and classification on resource-constrained microcontroller units (MCUs). • Extremely lightweight classification model for Halyomorpha Halys detection. • Joint improvement of hardware metrics without considerably compromising accuracy. [ABSTRACT FROM AUTHOR]

• A mechanical claw for seedlings fetching is designed to be suitable for low-damage seedling fetching methods. • The Unet network structure is improved, and the accuracy and robustness of the segmentation of the seedling leaves are improved. • A seedling point estimation algorithm is proposed to realize the calculation of seedling fetching posture and reduce seedling damage. • The seedling fetching effects of different seedling fetching methods were explored, and the effectiveness of the low-damage seedling fetching method was verified. Automatically, seedling fetching is a critical link in the automatic planting of facility agriculture. However, the damage caused by the mechanical claws to the seedling during the seedling fetching process seriously affects the survival rate of the seedling after being planted. To reduce the damage of the mechanical claw on the seedling leaves during seedling fetching, this paper proposes a low-loss seedling fetching method and designs a set of end-effectors suitable for the low-loss seedling fetching method. The low-damage seedling fetching way consists of a semantic segmentation algorithm and a seedling fetching point estimation algorithm, which can position the seedling leaves and calculate the optimal fetch seedling angle of the mechanical claw so that the mechanical claw can avoid seedling leaves during the seedling fetching to reduce the damage to the seedling. The backbone network of the Res-Unet model is composed of Resnet50, which can accurately extract the characteristics of plug seedlings. Comparative test results with other models show that our semantic segmentation model has high accuracy in positioning seedlings, with an MIoU value of 96.49%. Finally, a seedling fetching verification experiment was conducted, and the seedling damage rate was 4.16%. Compared with the ordinary seedling fetching method, the damage rate of the seedling was reduced by 15.28%. This method can minimize damage during the seedling fetching process and provide technical support for automated planting. [ABSTRACT FROM AUTHOR]

• Dual-branch block is designed to fuse the texture and color features. • Multi-scale and shrinkage blocks are proposed. • TC-MRSN improves identification in complex field background. • TC-MRSN outperforms than other state-of-the-art models. Maize diseases caused by fungal pathogens are the primary factor resulting in reduced maize yield. However, in practical complex background scenarios, diseases caused by spores, such as gray leaf spot and rust, usually exhibit characteristics including diverse propagation routes, similar lesion appearances at the initial stage of infection, and varying lesion sizes, which raise a challenging task to recognize similar diseases. Focusing on the accurate recognition of maize leaf diseases in complex backgrounds, this paper proposes a texture-color dual-branch multiscale residual shrinkage network (TC-MRSN) model based on deep learning. To preserve the characteristic information of small-sized lesions during the sampling process, texture feature extraction block and texture-color dual-branch block are designed to extract texture features from lesions and fuse them with RGB features. To reduce the interference of redundant background noise in the fusion feature, the multi-scale residual shrinkage module is presented to extract different receptive field features and process redundant noise through soft threshold. The proposed model is also deployed on mobile phones to enable real-time data collection and analysis. Detailed experimental and practical testing results show that TC-MRSN can achieve an average accuracy rate of 94.88% and 99.59% on complex background dataset and PlantVillage dataset, respectively, which is higher than those of the existing models ResNet50, VGG-ICNN, HCA-MFFNet by 5.2%, 2.5% and 1.8%, respectively. [ABSTRACT FROM AUTHOR]

• The first model based on GRU to reconstruct missing points in agricultural machinery trajectories is proposed. • A bidirectional feature extraction module is proposed to capture bidirectional information of the missing trajectory points. • A lightweight model is customized for massive agricultural machinery trajectory data. • The reconstruction deviation exceeds the Beidou GNSS positioning accuracy. Agricultural machinery trajectory data often encounters the phenomenon of missing trajectory points, and reconstructing these missing trajectory points is crucial for subsequent researches that require complete and high-quality agricultural machinery trajectory data. In this paper, we propose a new method called Fast-TRGRU to accurately and quickly reconstruct missing points in the trajectory of agricultural machinery. First, we design a novel data preprocessing process to improve the quality of raw data. Next, we introduce a bidirectional feature extraction module (BFE) to simultaneously utilize the bidirectional adjacent subtrajectory information of missing trajectory points. Then, we use gated recurrent unit (GRU) to extract spatiotemporal correlation features from time series data representing agricultural machinery trajectories to reconstruct trajectories. Finally, we reduce the training time of the model by adjusting the structure of the model in the GRU model. To verify the effectiveness of the model, we conducted experiments using 125 real agricultural machinery trajectory samples including 1,111,813 points provided by the Key Laboratory of Agricultural Machinery Monitoring and Big Data Applications of the Ministry of Agriculture and Rural Affairs. The experimental results show that when the trajectory missing rate is 10 %, the average reconstruction deviations of the model in longitude and latitude are only 1.80 m and 1.95 m, respectively. When the trajectory missing rate is 60 %, the average reconstruction deviation of the model in both longitude and latitude is still less than 4 m, which is higher than the positioning accuracy of the Beidou Navigation Satellite System. The Fast-TRGRU can provide complete and high-quality data for the analysis of agricultural machinery trajectories, effectively assisting in the research of optimizing agricultural machinery operation paths and evaluating agricultural machinery operation efficiency. [ABSTRACT FROM AUTHOR]

• Unmanned Aerial Vehicle (UAV) can detect rubber tree powdery mildew. • The proposed MSA-CNN model improves the classification performance. • Fused features perform better than single type of features. Rubber tree powdery mildew(PM) is one of the most critical leaf diseases of rubber trees. The epidemic of this disease can seriously affect natural rubber yields and necessitates timely monitoring, especially in the early stages. In recent years, unmanned aerial vehicle(UAV) hyperspectral imaging technology has been widely used in the field of crop disease identification. Therefore, this paper proposes a rubber tree PM detection method based on UAV low-altitude remote sensing and hyperspectral imaging technology. Firstly, spectral reflectance wavelengths, vegetation indices(VIs), and texture features(TFs) were extracted from the region of interest of the UAV hyperspectral image. Then, random frog(RFrog) was performed to select the optimal wavelengths(OWs), Pearson correlation coefficient(PCC) and sequence backward selection(SBS) algorithm to select the effective VIs and TFs. Secondly, the multi-scale selective attention convolutional neural network(MSA-CNN) model was constructed to detect PM based on OWs, VIs, TFs and their combinations. Moreover, the original images with spatial resolution of 10 cm were resampled to different spatial resolutions(20 cm , 40 cm , 60 cm , and 80 cm) to evaluate the effect of spatial resolution in PM monitoring. The results show that the proposed MSA-CNN model could sufficiently learn important features at different scales and obtain the best results in the full-wavelength dataset(OA = 93.79 %, Kappa = 92.16 %). Meanwhile, the models constructed using combining features(OWs + TFs, VIs + TFs, OWs + VIs + TFs) perform better than single features(OWs, VIs, TFs), and the highest performance was obtained for the OWs + VIs + TFs-based model(OA = 98.44 %, Kappa = 98.04 %). The optimal spatial resolution for PM monitoring was 10 cm. In addition, combining features could improve the classification accuracy of the early stages of PM. The results of the study provide a reference for accurate PM monitoring using UAV hyperspectral images. [ABSTRACT FROM AUTHOR]

• Pixel-wise facial landmark segmentation was done in cattle infrared images. • Ghost modules were introduced to reduce computational complexity. • ECA modules further improved segmentation performance. • Temperature parameters obtained on the segmentations were close to the ground truth. • Automated temperature collection will promote relevant uses of infrared thermography. In cattle, facial temperatures captured by infrared thermography provide useful information from physiological aspects for researchers and local practitioners. Traditional temperature collection requires massive manual operations on relevant software. Therefore, this paper aimed to propose a tool for automated temperature collection from cattle facial landmarks (i.e., eyes, muzzle, nostrils, ears, and horns). An improved UNet was designed by replacing the traditional convolutional layers in the decoder with Ghost modules and adding Efficient Channel Attention (ECA) modules. The improved model was trained on our open-source cattle infrared image dataset. The results show that Ghost modules reduced computational complexity and ECA modules further improved segmentation performance. The improved UNet outperformed other comparable models on the testing set, with the highest mean Intersection of Union of 80.76% and a slightly slower but still good inference speed of 32.7 frames per second. Further agreement analysis reveals small to negligible differences between the temperatures obtained automatically in the areas of eyes and ears and the ground truth. Collectively, this study demonstrates the capacity of the proposed method for automated facial temperature collection in cattle infrared images. Further modelling and correction with data collected in more complex conditions are required before it can be integrated into on-farm monitoring of animal health and welfare. [ABSTRACT FROM AUTHOR]

Quality grading and sorting are crucial post-harvest procedures to increase the market value of crops. To reduce the over-dependence on manual work for these tasks, automated sorting methods have been studied. However, we find existing research focuses mainly on round-like fruits and vegetables, and most machines can only grade one object at a time in their vision pipeline. To address these limitations, this paper proposes a comprehensive framework specifically for detecting and analyzing rod-like crops based on multi-object oriented detection. Zizania shoots are primarily used to validate our methods. To implement the deep learning models, an efficient oriented bounding box label software called OBBLabel is developed, and several large-scale image datasets are constructed. Both of the software and datasets are open-sourced for the community. Based on YOLOv8 architecture, the proposed YOLO-OBB model predicts oriented bounding boxes to extract all rod-like targets in image with 0.903 mAP@0.5. A multi-label recognition model YOLO-MLD then conducts quality grading and posture perception on individual target with 93.4% mean accuracy. Thus, precise position and quality information for all objects can be obtained in near real-time for subsequent suction cup-based sorting operations. Furthermore, according to the framework, this paper designs and manufactures a prototype sorting machine with edge intelligence for rod-like crops. • Framework for rod-like crops sorting with dataset, software, algorithms and machine. • Oriented detection model detects arbitrary-pose targets by oriented bounding boxes. • Multi-label recognition model parses the grade and posture of extracted single crop. • Machine with edge intelligence automates the workflow from sorting to packaging. • Annotation software and labeled datasets are open-sourced. [ABSTRACT FROM AUTHOR]

Crop counting is a crucial step in crop yield estimation. By counting, crop growth status can be accurately detected and adjusted, improving crop yield and quality. In recent years, with the rapid development of convolutional neural networks, deep learning-based object detection methods have been widely used in crop counting. By summarizing the research related to crop counting, this paper reviews the development status of object detection and crop counting. It then compares deep learning-based object detection counting methods with other counting methods. The paper also introduces public datasets and evaluation metrics commonly used for algorithmic models and provides a more in-depth analysis of the application of object detection in crop counting. Finally, the current problems that need to be solved, such as the lack of datasets, difficulties in small object counting, occlusion in complex environments, and some future directions are summarized. We hope this review will encourage more researchers to use deep-learning object detection methods in agriculture. • The paper compares the advantages and disadvantages of the methods currently used for crop counting tasks. • It is the first summary of object detection-based methods' application in crop counting tasks, covering commonly used models, issues, and challenges. • Public datasets utilized for crop counting tasks through object detection-based methods have been summarized. [ABSTRACT FROM AUTHOR]

• Recent advances in weak and no supervision learning and their applications in agriculture. • A new taxonomy of label-efficient algorithms was proposed according to the degree of required supervision. • The remaining challenges and potential future directions for label-efficient learning in agriculture were discussed. The past decade has witnessed many great successes of machine learning (ML) and deep learning (DL) applications in agricultural systems, including weed control, plant disease diagnosis, agricultural robotics, and precision livestock management. However, a notable limitation of these ML/DL models lies in their reliance on large-scale labeled datasets for training, with their performance closely tied to the quantity and quality of available labeled data. The process of collecting, processing, and labeling such datasets is both expensive and time-consuming, primarily due to escalating labor costs. This challenge has sparked substantial interest among researchers and practitioners in the development of label-efficient ML/DL methods tailored for agricultural applications. In fact, there are more than 50 papers on developing and applying deep-learning-based label-efficient techniques to address various agricultural problems since 2016, which motivates the authors to provide a timely and comprehensive review of recent label-efficient ML/DL methods in agricultural applications. To this end, a principled taxonomy is first developed to organize these methods according to the degree of supervision, including weak supervision (i.e., active learning and semi-/weakly- supervised learning), and no supervision (i.e., un-/self- supervised learning), supplemented by representative state-of-the-art label-efficient ML/DL methods. In addition, a systematic review of various agricultural applications exploiting these label-efficient algorithms, such as precision agriculture, plant phenotyping, and postharvest quality assessment, is presented. Finally, the current problems and challenges are discussed, as well as future research directions. A well-classified paper list that will be actively updated can be accessed at https://github.com/DongChen06/Label-efficient-in-Agriculture. [ABSTRACT FROM AUTHOR]

• Improved MC-AlexNet structure is established for opening state recognition. • Binocular grayvalue deformation flower template matching method is designed. • Another 3D position and pose recognition method is established. It is necessary to recognize the tomato pollination features for the designing demand of intelligent and precise tomato supplementary pollination equipment. Mentioned pollination features include flower opening state and the three-dimensional position and pose of flower anther. Tomato flower pollination features recognition method is designed in this paper based on the deep learning full opened flower recognition model and binocular template matching three-dimensional information recognition method. First of all, a binocular stereo vision system is built to acquire the tomato flower images in greenhouse with natural lighting. Which can help vision system avoid the impact of inconsistent light intensity on image recognition. The acquired images were equalized with three groups parameters to labeled images. And the improved MC-AlexNet deep learning model is established to recognize the full opened tomato flowers in the left image of image pair acquired with binocular vision system. Then, the template is created with the full opened flower recognition result of deep learning model in left image based on gray value and deformation template matching method. And the template matching is established to recognize the corresponding full opened flowers in the right image of image pair. Finally, with the template matching result, the anther segmentation is conducted to calculate three-dimensional position and pose of anther. The experimental results show that the accuracy of full opened tomato flowers recognition model is 96.23%. The average position recognition deviation of anther is 5.94 mm. And the average anther pose recognition deviation in three plane is 6.24° for the images that anthers can be observed. And the average time consuming is about 143.18 ms per image pair. It can be concluded that the method of binocular template matching established in this paper can fulfill the demand of supplementary pollination equipment design, and the research result lays a foundation for designing and improvement of tomato supplementary pollination equipment in greenhouse. [ABSTRACT FROM AUTHOR]

• Knowledge graph theory was applied to extract and compile information from domain textual resources. • Deep-learning models were applied to extract entities and relationships from dairy literature. • The use of domain dictionaries can accelerate the named entity recognition corpus preparation process. • Natural language could be parsed into database query language by semantic parsing methods to directly retrieve information from the database. Three weeks prior to calving to three weeks after calving, the transition period poses challenges for dairy cattle and farmers. Vast changes in housing, feeding, and reproduction might result in milk drop, metabolic and reproductive diseases. Moreover, most of the metabolic processes are intricately linked as many conditions can coexist. This challenge means that dairy producers and their advisors have difficulty drawing concise conclusions because of all aspects and relationships in transition cow management. Herein, machine-learning techniques and knowledge-graph theory were explored with a view to creating a decision-support system that could provide producers and their advisors with knowledge from domain literature. Specifically, knowledge is modelled as entities and relationships in knowledge graph theory, and natural language models were developed to extract information as knowledge graphs. A dataset comprising 1152 sentences from 20 papers was created and split into 922 sentences for training and 230 sentences for testing. Sequentially, two deep learning models were trained to extract entities and relationships respectively. For training results, a Bi-directional Long-Short-Term Memory model was applied for the entity extraction task and obtained an F1 score of 80 %. As for relationship extraction, a Transformer-based model was deployed but yielded a low F1 of 23 %, thus another pre-trained Transformer model with 89 % accuracy was deployed into the system. After feeding the domain literature into the deep-learning models, a knowledge graph of 1,576 nodes and 3,456 edges was constructed and stored in the graph database Neo4j. Afterward, a semantic parsing method was used to allow users to conduct question answering through the knowledge graph in natural language. In addition, to determine the quality of answers that the knowledge built from the papers, answers were sampled and evaluated based on human judgment. On average, answers scored 7.5 out of 10 and proved informative with respect to the original literature. Although the final interactive results demonstrated a high degree of visualization and scalability, this study primarily sought to demonstrate its feasibility. For tailored commercial applications, further improvements could be implemented in knowledge graph expansion and reasoning. [ABSTRACT FROM AUTHOR]

[Display omitted] • This paper proposed a new method for evaluating BCS based on shape analysis. • This paper provided a highly automatic GUI tool to evaluate the BCS of livestock. • The method in this paper is more general and is applicable to multiple species. • The method in this paper can be extended to other shape related research. Body condition has always been an important indicator reflecting the health status of livestock, and new automated and contactless methods for assessing it have also been continuously proposed. In recent years, with the continuous expansion of research on quantitative analysis of three-dimensional shapes in the field of agriculture, amount of studies has shown that quantitative analysis of shapes can promote the development of smart agriculture. This article proposed a method for quantitatively analyzing the local three-dimensional shape of livestock to evaluate their body conditions. This is a universal method that can be applied to various species. It uses the method of point-to-point correspondence in the three-dimensional shape to finely calculate the difference between shapes, and then maps the difference value to the range of the body condition score. The article analyzed the shape of 198 dairy cows, 100 beef cattle, and 201 pigs using body condition scores as test parameters. The accuracy of body condition evaluation within the range of error of 0.5 for dairy cows, 1 for beef cattle, and 0.5 for pigs are 100%, 87.61%, and 92.09%, respectively. Compared to previous methods for assessing body condition, the method proposed in this paper is more accurate and versatile. The dataset used in this study, along with the code implementation have been made publicly available in https://gitee.com/kznd/lshape-analyser/tree/master/MATLAB. [ABSTRACT FROM AUTHOR]

• This paper presents a novel one-shot domain adaptive real-time 3D detection method that addresses the challenge of insufficient target domain samples by designing a semantic-geometry-intensity space generator. This generator learns both intra-category similarity and inter-category difference. • The integration of semantic-geometry-intensity space-based classifier and centerpoint-based anchor-free locator is proposed to complete the feature-category mapping and determine 3D bounding boxes. • The switching between object sample enhancer and fusion point cloud generator is proposed to fully learn features of each category and rectify the attitude deviation. By introducing deep learning, LiDAR-based solutions have achieved impressive accuracy in 3D obstacle detection. However, gathering and labeling sufficient samples is the precondition for the effectiveness of existing solutions. This precondition is difficult to satisfy in actual farmland due to the scarcity of distinctive obstacle samples as well as the time-consuming and specialized labeling process. In practice, detection models trained on specific datasets may fail to generalize well to real farmland, as they lack adaptability to different categories and scenes. To address this limitation, this paper proposes a novel one-shot domain adaptive real-time 3D obstacle detection method based on semantic-geometry-intensity fusion strategy. By introducing the concept of one-shot domain adaptation, the proposed method enables fine-grained 3D obstacle detection with just one sample per category. Specifically, a semantic-geometry-intensity space generator is designed to bridge the category gap between training and test samples. The integration of semantic-geometry-intensity space-based classifier and centerpoint-based anchor-free locator is designed to keep a decent balance between accuracy and efficiency. The switching between object sample enhancer and fusion point cloud generator is designed to handle distribution differences of both points and categories. The obstacle detection system designed based on the proposed method has been tested in real farmland, achieving an overall F 1 score of 89.54% and a Frame Rate of 21.32 Frames Per Second (FPS). These experimental results demonstrate the high accuracy and efficiency of the proposed method in performing obstacle detection. [ABSTRACT FROM AUTHOR]

• Proposed universal digital twin architecture for dairy cows. • Implementation of a low-cost, easy-to-implement digital twin solution for dairy cows. • Development of a new collar, anchor point, and software, demonstrated through effective feeding behavior classification. • Accuracy of recognition for region-specific behavior improved by combining indoor positioning data with IMU data. The digital twin of cows holds significant promise for advancing animal welfare and production efficiency. This paper aims to propose an architecture for digital twins of cows that covers their entire lifecycle. A digital twin solution has been developed that utilizes indoor positioning data and inertial measurement unit (IMU) data to construct a cow's digital shadow. As an example, the paper utilizes the classification of cow feeding and non-feeding behaviors to study the digital twin perception and modeling methods. A custom-made collar integrated with utilized ultra-wideband (UWB) chips and inertial measurement units (IMUs) was utilized to collect real-time location and neck movement data from five healthy non-lactating Holstein cows. The collected data was transmitted via UWB signals to the positioning anchor and subsequently forwarded to a local server. To classify the feeding and non-feeding behaviors of the cows, three methods were employed: Support Vector Machines (SVM), K-Nearest Neighbor (KNN), and Long Short-Term Memory (LSTM). According to the experimental results, all three classification methods were effective, however, LSTM outperformed the others. Employing solely IMU data and implementing the LSTM, the precision of identifying bovine foraging behavior reached 91.05%, with concomitant precision and recall rates of 92.23 and 91.35%, respectively. Through an integration of the data from indoor position detection and IMU devices and the employment of LSTM, the accuracy of identification increased to 94.97%, with a precision rate of 99.99% and a recall rate of 93.86%. The trial of the digital twin solution demonstrated the rationality and technical feasibility of the digital twin architecture, which holds significant reference value for the development of animal digital twins in the animal husbandry industry. [ABSTRACT FROM AUTHOR]

• An adaptive cruise weeding robot for paddy fields was developed for the first time. • A real-time rice seedling recognition method based on MW-YOLOv5s was proposed. • A navigation line fitting method for paddy field adaptive operations was proposed. Weed control in paddy fields is a critical agronomic practice for enhancing rice yield, in which mechanical weed control is widely used due to its high weed control rate and convenient operation. However, traditional mechanical weed control methods require manual operation, leading to increased operational costs. Therefore, adaptive cruise weeding robots hold significant application promise and market potential. In this paper, an adaptive cruise weeding robot for paddy fields based on improved YOLOv5 is designed to adaptively perform weeding operations. Firstly, a real-time rice seedling recognition method based on MW-YOLOv5s is proposed by establishing a rice seedling dataset under different paddy field environments and growth stages. The model combines the Backbone network structure with MobileViTv3, replacing GIoU_loss with WIoU_loss, which significantly improves the precision and speed of rice seedling recognition. Next, the MW-YOLOv5s was integrated into a paddy weeding machine, and seedling navigation lines were extracted using the least square method. Finally, the control system autonomously operates the weeding machine in real time through feedback control based on the navigation path. The test results showed that the MW-YOLOv5s model exhibits strong recognition performance for rice seedlings under different paddy field backgrounds and growth stages, with a precision of 90.05 % and an mAP of 92.32 %, while the real-time performance reaches 19.51 FPS, which meet the requirements for real-time operation of the weeding machine in paddy fields. The results of the paddy field weed control experiment reveal a weed control rate of 82.4 % and a seedling injury rate of 2.8 %, which meet the agronomic requirements for mechanical weed control in paddy fields. The research findings contribute to the design of intelligent agricultural equipment and promote the practice and application of machine vision in the field of paddy field plant protection. [ABSTRACT FROM AUTHOR]

Plant disease is one of the major problems in agriculture. Diseases damage plants, reduce yields and lower the quality of the produce. Traditional approaches to detecting plant diseases are usually based on visual inspection and laboratory testing, which can be expensive and time-consuming. They require trained plant pathologists as well as specialised equipment. Several studies demonstrate that artificial intelligence (AI) methods can produce promising results. However, AI methods are generally data-hungry and require large annotated datasets, and the collection and annotation of such datasets can be a limiting factor. It often appears that only a small amount of data is available for certain disease types. Whereas the performance of typical AI methods drops significantly when they are trained with inadequate data. This paper proposes a novel few-shot learning (FSL) method to detect plant diseases and alleviate the data scarcity problem. The proposed method uses as few as five images per class in the machine learning process. Our method is based on a state-of-the-art FSL pipeline called pre-training, meta-learning, and fine-tuning (PMF), integrated with a novel feature attention (FA) module; we call the overall method PMF+FA. The FA module emphasises the discriminative parts in the image and reduces the impact of complicated backgrounds and undesired objects. We used ResNet50 and Vision Transformers (ViT) as the feature learner. Two publicly available plant disease datasets were repurposed to meet the FSL requirements. We thoroughly evaluated the proposed method on the PlantDoc dataset, which contains disease samples in field environments with complex backgrounds and unwanted objects. The PMF+FA method with ViT achieved an average accuracy of 90.12% in disease recognition. The results demonstrate that the PMF+FA pipeline consistently outperforms the baseline PMF. The results also highlight that the method using ViT generates better results than ResNet50 for diagnosing complex data. ViT and ResNet50 implementations are computationally efficient, taking 1.11 and 0.57 ms on average per image to evaluate the test set respectively. The high throughput and high-quality performance with only a small training dataset indicate that the proposed technique can be used for real-time disease detection in digital farming systems. • A few-shot learning approach is proposed for detecting plant diseases from images. • High performance was achieved using only a few training images per disease. • Vision transformers further improve performance, especially on complex samples. • Our proposed attention module reduces the impact of complicated backgrounds. • The high throughput and performance indicate its potential in precision farming. [ABSTRACT FROM AUTHOR]

• We provide a UAV-based multispectral imagery approach objectifying FHM strategies. • A CNN was trained using 17,643 images of trees with health status and species labels. • Analysis was done at 235 plots distributed across Bavaria over three years. • The highly heterogenous dataset benefits model generalization and transferability. • We provide a pipeline to generate reference data and for DL image classification. Uncrewed Aerial Vehicles (UAVs) have emerged as a promising tool for complementing terrestrial surveys, offering unique advantages for forest health monitoring (FHM). UAVs have the potential to improve or even replace core tasks such as crown condition assessment, bridging the gap between ground-based surveys and traditional remote sensing platforms. However, present approaches have not yet fully exploited the very high temporal resolution and flexible and convenient utilization that UAVs offer even under cloudy skies. In this paper, we provide a standardized data pipeline to semi-automatically generate reference data and for monitoring forest health by merging ground-based and UAV-based data related to species-specific forest health. Furthermore, we investigated the potential of Convolutional Neural Networks (CNNs) to classify the main tree species and their crown conditions based on the reference data. Therefore, we acquired high resolution multispectral drone imagery of 235 different ICP large scale forest monitoring plots (Level-I plots) distributed across Bavaria for three consecutive years (2020–2022). Using this highly heterogeneous time-series dataset, encompassing diverse weather and lighting conditions, forest stand characteristics, and spatial distribution of study areas, we successfully classified five tree species, three genus level classes and dead trees, including the health status of the main tree species occurring in Germany. This way we managed to classify 14 distinct classes with an average macro F1-score of 0.61 using the EfficientNet CNN architecture. The highest class-specific F1-score apart from the class of dead trees (0.97) was achieved by the class of Picea abies healthy (0.80). If participating countries of the ICP Forests program adopt our approach to harmonize terrestrial and UAV-based monitoring, many ground-based tasks could be reduced or replaced, leading to significant time and cost savings. We provide standardized and open-source monitoring and analysis strategies that can be potentially extended throughout Europe. Our findings demonstrate that UAV monitoring and deep learning can modernize forest management for efficiency and sustainability. We recommend integrating drones with ground surveys in forest monitoring systems to take advantage of their benefits. [ABSTRACT FROM AUTHOR]

• Detailed review of over 300 research papers. • Object Detection and Tracking are suitable techniques to enhance Precision Farming. • Critical discussion regarding the challenges of Object Detection and Tracking. • Potential research lines to improve Object Detection and Tracking are identified. Object Detection and Tracking have gained importance in recent years because of the great advances in image and video analysis techniques and the accurate results these technologies are producing. Moreover, they have successfully been applied to multiple fields, including the agricultural domain since they offer real-time monitoring of the status of the crops and animals while counting how many are present within a field/barn. This review aims to review the current literature on Object Detection and Tracking within the field of Precision Farming. For that, over 300 research articles were explored, from which 150 articles from the last five years were systematically reviewed and analysed regarding the algorithms they implemented, the domain they belong to, the difficulties they faced, and which limitations should be tackled in the future. Lastly, it examines potential issues that this approach might have, for instance, the lack of open-source datasets with labelled data. The findings of this study indicate that Object Detection and Tracking are critical techniques to enhance Precision Farming and pave the way for robotization for the agricultural sector since they provide accurate results and insights on crop and animal management, and optimize resource allocation. Future work should focus on the optimal acquisition of the datasets prior to Object Detection and Tracking, along with the consideration of the biophysical environment of the farming scenarios. [ABSTRACT FROM AUTHOR]

Effective identification of tomato plant traits is crucial for timely monitoring and evaluating their growth and harvest. However, conducting stress experiments on multiple tomato genotypes introduces challenges due to the nature of the data. One of these challenges arises from an imbalanced sample distribution, potentially leading to misclassification between classes and disruptions in model recognition. This paper addresses the effect of these challenges by considering the imbalanced classes of flowers, fruits, and nodes and proposing an improved detection approach through data balancing. A novel data-balancing approach is introduced in this study to overcome the issue of imbalanced data. The proposed solution involves the implementation of a YOLOv8 deep learning model, which effectively detects flowers, fruits, and nodes in tomato plants. This model significantly enhances the ability of the algorithm to detect objects of varying sizes within complex environments. To further bolster the recognition capability of the targeted classes, the proposed model integrates a Squeeze-and-Excitation (SE) block attention module into its head architecture. This module strengthens the model recognition ability by giving increased attention to the studied classes, thereby enhancing overall detection performance. The results demonstrate that the data balancing approach successfully improves the model performance in response to the data challenges. When applying the technique of pre-training the optimal weights obtained from balanced data on imbalanced data, the SE-block module showed significant improvements in outcomes. • YOLOv8-based single-stage detector for tomato plant phenotyping. • Data balancing of nodes, fruit, and flowers on a challenging tomato images dataset. • Impact of attention mechanisms in the localization of small phenotypical traits. • Performance comparison of several optimization strategies, preprocessing methods, and stand-alone detectors. [ABSTRACT FROM AUTHOR]

• We introduce a unique dataset of UAV thermal imagery, meticulously collected under diverse heights, angles, and climatic conditions, designed to challenge and validate thermal imaging analysis. • An approach for tracking mammalian herbivores in thermal imagery using the Multi-Object Tracking and Segmentation model Track R-CNN is proposed. • 3D convolutions outperform LSTM convolutions as a temporal component. • The association head outperforms optical flow as a tracking mechanism. • The proposed approach is robust across a wide range of lighting and thermal conditions and on imagery collected at various heights and angles. Lightweight Unmanned Aerial Vehicles (UAVs) are emerging as a remote sensing survey tool for animal monitoring in several fields, such as precision livestock farming. Together with state-of-the-art computer vision techniques, UAV technology has drastically escalated our ability to acquire and analyse visual data in the field, lowering both costs and complications associated with collection and analysis. This paper addresses monitoring mammalian herbivores using the unexploited field of thermal Multi-Object Tracking and Segmentation (MOTS) in UAV imagery. In our research, a state-of-the-art MOTS algorithm (Track R-CNN) was trained and evaluated in the segmentation, detection and tracking of dairy cattle. Data collection was carried out in two farms with a UAV carrying a thermal camera at various angles and heights, and under different light (overcast/sunny) and thermal (16.5 °C range) conditions. Our findings suggest that dataset diversity and balance, especially regarding the range of conditions under which the data was collected, can significantly enhance tracking efficiency in specific scenarios. For training the algorithm, transfer learning was used as a knowledge migration method. The performance of our best model (68.5 sMOTSA, 79.6 MOTSA, 41 IDS, 100 % counting accuracy, and 87.2 MOTSP), which utilizes 3D convolutions and an association head, demonstrates the applicability and optimal performance of Track R-CNN in detecting, tracking, and counting herbivores in UAV thermal imagery under heterogenous conditions. Our findings demonstrate that 3D convolutions outperform Long-short Term Memory (LSTM) convolutions. However, LSTM convolutions also show optimal performance, offering a viable alternative. Furthermore, our results highlight the inability of Optical Flow to track motionless animals (-15 sMOTSA, −4.1 MOTSA and 2076 IDS) and the proficiency of the association head in differentiating static animals from the background. This research contributes to the growing body of knowledge in automated mammalian herbivore monitoring, with potential applications such as precision livestock farming and wildlife conservation. [ABSTRACT FROM AUTHOR]

The scientific progress in artificial intelligence and robotics has enabled precision viticulture to pursue sustainability and improve the final yield. For instance, monitoring the canopy volume of each plant can allow the correct ripening of the bunches. In this context, this paper proposes a novel approach for the characterization of biomass volume using images acquired in a vineyard with the low-cost Azure Kinect RGB-D camera. Semantic image segmentation is implemented using three encoder–decoder deep architectures (U-Net, DeepLabV3+, and MANet) to produce accurate masks of the vine leaf structure. In a transfer learning approach, a public dataset acquired with the Intel RealSense D435 depth camera is used to train the segmentation networks. Then, a complete pipeline to estimate possible changes in biomass volume is presented. Experiments are run to analyze the biomass removed during the trimming process of grapevine plants. The best segmentation result is obtained by the U-Net architecture with ResNet50 backbone, showing an accuracy of 92.10%, although the training and test sets consist of images acquired by different cameras. However, the DeepLabV3+ network with ResNeXt50 backbone, which scores an accuracy of 90.25% on the test set, gives the best estimate of the removed biomass, requiring the shortest time for training. These outcomes prove the potential capability of this automatic approach for controlling leaf growth and ensuring sustainable viticulture practices. • Low-cost RGB-D cameras for Biomass characterization before and after the trimming process. • Deep learning-based semantic segmentation networks designed for high generalization. • Combination of multimodal data (color and depth) to extract only significant features. • Registration of point clouds to allow the exact characterization of the removed biomass. [ABSTRACT FROM AUTHOR]

• An anti-lodging machine, comprising an anti-lodging device and an AGV (Automated Guided Vehicle) was designed to replace manual methods. • A machine vision system was developed to aim the anti-lodging device at the crown buds. • The YOLOv5s algorithm is enhanced for detecting immature pineapple crown buds. • The mAP_0.5 of the improved YOLOv5s algorithm is 3.71 %, 1.69 %, and 1.19 % higher than the original YOLOv5s, SSD, and Faster R-CNN, respectively. • The operating speed of this intelligent anti-lodging method is 2.7 times faster than manual binding. Sunscorch, a common disease in pineapple cultivation, is caused by direct sunlight on the surface of lodging pineapple fruits. To mitigate this issue, pineapple fruits are often bagged or supported with fixing rods. This paper presents the design of an anti-lodging machine, comprising an anti-lodging device and an AGV (Automated Guided Vehicle), and develops a machine vision system to precisely aim the anti-lodging device at the crown buds. The YOLOv5s algorithm is enhanced for detecting crown buds. The improvement of the YOLOv5s algorithm includes the introduction of the attention mechanism CBAM to enhance feature extraction, the addition of BiFPN and GhostNet to strengthen feature fusion and to reduce detection time and computational cost. By transforming pixel coordinates, the relative coordinates between the crown bud and the device are calculated to achieve accurate targeting. The experiments demonstrate that the precision, recall, and map_0.5 of the improved YOLOv5s algorithm are 91.89 %, 92.25 %, 94.01 %, and the detection speed is 15.71 ms/image. The average detection speed is 1.14 ms/image and 9.45 ms/image faster than that of SSD and Faster R-CNN, benefiting from a parameter number only 3/4 and 3/7 of SSD and Faster R-CNN, respectively. The results of contrast experiments indicate that, although the lodging prevention rate of this intelligent anti-lodging method is 4 % lower than the manual binding, its operating speed is 2.7 times faster. [ABSTRACT FROM AUTHOR]

Pine Wilt Disease (PWD) is a devastating disease that affects forest ecosystems and has no known cure. Early detection is essential for suppressing infections of PWD. Recent efforts have focused on leveraging deep learning-based algorithms for the early detection of PWD. However, data collection, labeling, and quality assurance for such a method are costly and challenging. Particularly for PWD, the data collection period is limited, restricting the diversity of the dataset. To overcome these challenges, this paper introduces a virtual forest considering PWD created using 3D rendering tools, from which we built a synthetic dataset. Furthermore, to ensure the resemblance of the synthetic data to the real data, we employed Image-to-Image (I2I) translation techniques. We used the EfficientNetv2-S model to compare the results from each dataset. We were able to confirm the potential of our model trained solely on the PWD synthetic dataset for real PWD detection. Moreover, the model trained on an ensemble comprising both real and synthetic data exhibited improved performance, achieving an F1 Score of 92.88%. Using the I2I translation technique for an ensemble of real and synthetic data also demonstrated enhanced performance. This result confirms the validity and practicality of the synthetic data proposed in this study. The findings of this study are expected to contribute to the broader forest ecosystem preservation and agricultural management by extending to other forest disease detection and agricultural fields. Details of our dataset and code are available at https://github.com/dydgns2017/PWD-Synthetic-Dataset. • Pine Wilt Disease requires early detection due to its severity and lack of cure. • Our synthetic dataset creation outperforms traditional PWD data collection process. • Real and synthetic data combination improves PWD F1 Score to 92.88%. • Synthetic data method aids forest preservation, applies to other agriculture. [ABSTRACT FROM AUTHOR]

• Cows and their abdomen areas were located using YOLOv5. • Individual cows were measured remotely, non-invasively, and simultaneously. • Heat stress was detected when the respiration rate reached 60 breaths per minute. In cattle, respiration rate (RR) provides researchers and practitioners with valuable physiological information. However, traditional visual observation requires massive labour and is impractical for large-scale commercial farms. Recently developed vision-based RR measurement methods are highly limited to measuring a small number of animals in a controlled environment. Therefore, this paper aimed to propose a vision-based multi-object RR measurement method for dairy cows lying in free stalls. An RGB camera was aimed at the lying zone and was able to cover about 16 stalls. The proposed framework first utilised two YOLOv5-based networks to segment cow instances and detect cow flank objects. Next, an object tracker was used to link the predictions of each cow throughout the video clip. The Lucas-Kanade optical flow was then calculated specifically on the overlapped area of the cow mask and the flank bounding box. Finally, RR was extracted using Fast Fourier Transform. The results show that the proposed method had a precise RR measurement with a correlation coefficient of 0.944, a root mean square error of 5.35 breaths per minute, and an intraclass correlation coefficient of 0.974 when compared to visual observation. The piecewise regression models identified a change in RR when the ambient temperature reached 23.6 °C or the temperature-humidity index reached 72. The corresponding RR thresholds were 60.9 and 60.2 breaths per minute, respectively. Collectively, these results can be used to inform an automated local cooling system, e.g., fans in the lying area. However, more experiments and calibration with data collected using more cost-effective video recording systems are required before this technology can be applied on farms. [ABSTRACT FROM AUTHOR]

• Amaranthus palmeri growth stage recognition was possible with YOLO architectures. • YOLO v8-X achieved the highest mAP@[0.5:0.95] for growth stage recognition of 47.34% • A. palmeri single-class outperformed 8-class recognition by over 20% (v7-original) • The first high-quality, benchmarking, growth stage dataset is provided. Many advanced image-based precision agricultural technologies for plant breeding, field crop research, and site-specific crop management hinge on the reliable detection and phenotyping of plants across highly variable morphological growth stages. Convolutional neural networks (CNNs) have shown promise for image-based plant phenotyping and weed recognition, but the ability to recognize growth stages, often with stark differences in appearance, is uncertain. Palmer amaranth (Amaranthus palmeri) is a particularly challenging weed plant in cotton (Gossypium hirsutum) production; due to high genetic diversity, it exhibits highly variable plant morphology across growth stages over a growing season, as well as between plants at a given growth stage. This paper investigates eight-class growth stage recognition of A. palmeri in cotton as a challenging detection case study for You Only Look Once (YOLO) architectures. In total, 26 architecture variants from YOLO v3, v5, v6, v6 3.0, v7, and v8 are compared on an eight-class growth stage dataset of A. palmeri. The highest mean average precision (mAP@[0.5:0.95]) for recognition of all growth stage classes was 47.34% achieved by v8-X, with inter-class confusion across visually similar classes. With all growth stages grouped as a single class, performance increased, with a maximum mAP@[0.5:0.95] of 67.05% achieved by v7-Original. Single-class recall of up to 81.42% was achieved by v5-X, and precision of up to 89.72% was achieved by v8-X. Class activation maps (CAM) were used to understand model attention on the complex dataset. Fewer classes, grouped by visual or size features improved performance over the ground-truth eight-class dataset. Successful growth stage detection highlights the substantial opportunity for improving plant phenotyping and weed recognition technologies with open-source object detection architectures. Additionally, the first open-access, benchmarking growth stage dataset ' Palmer amaranth growth stages – 8 (PAGS8) ' is presented, which is available here: https://weed-ai.sydney.edu.au/datasets/5c78d067-8750-4803-9cbe-57df8fae55e4. [ABSTRACT FROM AUTHOR]

• A transformer-based multiple fish tracking network: TFMFT. • Multiple Association (MA) method for maximum matching. • Empirical studies on diverse transformer tracking architectures. • A multiple fish tracking dataset with various fish species and scenarios. Recent advancements in fish tracking methodologies provide valuable solutions for assessing fish growth, marine fisheries, and biological research. In particular, there has been a burgeoning interest in vision-based methods for fish tracking, owing to the enhanced computational capabilities facilitated by deep learning models. However, these methods face several challenges, including poor fish detection performance under complex backgrounds, the potential for identification switches caused by the non-rigid features and occlusions of fish, and the limited fault tolerance of extant approaches. In this paper, a t rans f ormer-based m ultiple f ish t racking model (TFMFT) is proposed, specifically designed to address the issue of instance loss of fish targets in aquaculture ponds with complex background disturbance. In particular, we introduce a Multiple Association (MA) method that bolsters fault tolerance in tracking by synthesizing simple Intersection-over-Union matching in the identification (ID) matching module. Through empirical studies across diverse Transformer-based models, we comprehensively assessed the influence of architecture design on data requirements. Furthermore, to evaluate the performance and generalizability of fish tracking models, we present the Multiple_Fish_Tracking_2022 (MFT22) dataset. The results demonstrate that TFMFT achieves 30.6% IDF1 (Identification F-Score) on the MFT22 dataset, outperforming the state-of-the-art by 10.9% and showcasing superior performance over other models. The resources and pre-trained model will be available at: https://github.com/vranlee/TFMFT. [ABSTRACT FROM AUTHOR]

Automatic fruit blossom detection plays a crucial role in agricultural intelligence to predict fruit yield. Existing deep learning methods for vision-based fruit blossom detection oftentimes rely on large labeled samples and are tailor for a single category of fruit blossoms, limiting their flexibility and adaptability in real applications. In this paper, a fast and data-efficient framework is proposed to achieve multi-class fruit blossom detection using few labeled samples, which not only enhances the flexibility of model training by reducing the need for extensive samples but also extends the model's applicability to a wider range of blossom categories. Specifically, the proposed framework incorporates the paradigm of few-shot object detection into a lightweight two-stage detector named CenterNet2. To improve the recall of blossom proposals, we introduce a location guidance module (LGM) to highlight foreground regions in query images that bear resemblance to the blossom reference provided by support images. Moreover, a contrastive learning scheme (CLS) is introduced to further distinguish fruit blossom categories with similar appearances, enhancing classification accuracy. Experimental results demonstrated that our framework achieved comparable performance to existing state-of-the-art models, with a mean average precision of 74.33% over 12 categories of fruit blossoms and a frame per second rate of 47. In summary, our proposed framework can reduce data dependency and enhance flexibility of existing deep learning detectors, making it a promising solution for automatic monitoring in real-world agricultural applications. • The proposed framework can detect multi-class fruit blossoms with few labeled samples. • A location guidance module is introduced to guide the framework to produce more accurate blossom proposals. • A contrastive learning scheme is introduced to distinguish between different fruit blossoms with similar appearances. • The proposed framework achieves comparable performance to existing state-of-the-art models and faster inference speed. • The proposed framework can reduce data dependency and enhance flexibility of existing deep learning detectors. [ABSTRACT FROM AUTHOR]

• A staged data preprocessing strategy for honey peach young fruit is designed. • The SCAE is proposed to ensure the discriminability of the extracted features. • The SCAE is embedded into the DeepSORT to form the DeepSORT-SCAE approach. • The DeepSORT-SCAE can measure the spatially continuous trajectory of the young fruit. • The effectiveness of the developed method is verified by experiments. The young fruit bagging robot usually preferentially selects the object with largest prediction box size for bagging. However, due to the uncertainty in the output of the traditional models, the expected endpoint of mechanical motion may change constantly, which affects the bagging efficiency. This paper proposes a novel tracking algorithm for honey peach young fruit targets. Firstly, according to the proportion of young fruit prediction box in the original image, a staged data preprocessing strategy is designed. Then, the siamese convolution autoencoder (SCAE) is developed by combining autoencoding mechanism with siamese training mode, and the SCAE is trained in two-stages to ensure the discriminability of the extracted appearance features. Finally, the SCAE is inserted into the DeepSORT architecture to generate the DeepSORT-SCAE object tracking algorithm, which can measure the spatially continuous trajectory of multiple young fruit targets within video streams. The effectiveness of the developed method is verified by experiments. [ABSTRACT FROM AUTHOR]

• A dataset on small target pests on greenhouse sticky traps. • A novel solution was developed for counting multiple small targets based on density maps, simultaneously. • An adaptive sample equalization algorithm was developed to tackle the issue of class imbalance. Whiteflies and fruit flies are common pests that adversely affect greenhouse crops, so it is vital to control their numbers promptly. Current research involves setting up artificial monitoring devices with fixed yellow sticky traps (YSTs), for pest surveillance. However, these devices can only be installed in specific environments and additional interventions may impact the actual distribution of greenhouse pests. On the other hand, detection methods, which are the most commonly used, struggle to accurately detect overlapping or closely interacting small insects, which is a common occurrence in the presence of pests. In this paper, we propose a novel solution for the multi-species pest recognition and counting method based on a density map to tackle pest counting in the greenhouse. To reduce interference between different types of objects, we used two annotation maps to represent the distribution of each object separately. Then, two counting branches are designed to count the pests, respectively. Finally, the detected pests are mapped onto the original image to display their spatial positions. In addition, an adaptive sample equalization method is proposed to address the issue of class imbalance in images, and the atrous spatial pyramid pooling (ASPP) structure and the CBAM attention mechanism are added to the fruit fly branches to focus on learning fruit fly features from crowded edges of YSTs. When applied to split images, our method achieved reliable counting performance in counting both whiteflies and fruit flies, with R2 scores of 0.973 and 0.972, respectively. This method has been embedded into mobile devices to assist greenhouse workers in monitoring pest populations in the field. The experiments show that the proposed method provides a promising direction for counting multiple small target insects and monitoring pest populations in crowded images. At the same time, it demonstrates the potential of density map methods in various small target detection scenarios. [ABSTRACT FROM AUTHOR]

• New real-time tech for automatic berry counting aids farmers in thinning. • Novel method to predict berry number from a single 2D image. • Design and implementation of 8 key features from a compact deep learning model. • Achieves low mean absolute error (MAE) of 2.60 berries on extensive dataset. The advent of smart agriculture has revolutionized and streamlined various manual tasks in grape cultivation, one of which is berry thinning. This task necessitates experienced farmers to selectively remove a specific number of berries from the working bunch, as guided by the remaining number of berries in the bunch. In response, this paper introduces a novel real-time edge computing application that automates the process of counting the berries in a working bunch using a single 2D image. The proposed application employs YOLOv5-based object detection techniques (Jocher, 2021) to distinguish each working bunch and the visible and slightly occluded berries contained therein. The key contribution of this paper is to accurately predict the number of berries in the whole bunches including those not visible in a 2D image by harnessing the output from object detection to devise features based solely on bounding box information. In addition, the feature set is optimized by employing a wrapper feature selection method (Kohavi & John, 1997), in consideration of the limitations of edge computing devices. The eight selected features yield a mean absolute error (MAE) of 2.60 berries, tested on a dataset of 26,230 images. Only a slight increase over the initial 19-feature set, which achieved an MAE of 2.42 berries. Furthermore, the proposed approach has been successfully implemented and tested on an Android smartphone, the Sony Xperia 1 III, without the need for an internet connection. The overall computation time per image stands at an average of 0.333 s, confirming its potential for real-world application. [ABSTRACT FROM AUTHOR]

• This study proposed detection and localization methods based on deep learning and active sensing for harvesting robots in real-world environments with occlusion and varying lighting conditions. • To solve the picking problem caused by the peduncle's random tilting, this study proposed a method to calculate the peduncle's inclination angle for controlling the end-effector to make the corresponding rotation. • Experimental results indicated the error between the estimated occlusion ratio and the genuine value is 16% in average, and the active sensing method improved the confidence score in occluded situations by over 50%. And the overall successful picking rate of 300 trials is 90%. • The proposed active methods have a 33% increase in precision and a 43% increase in efficiency compared to constant methods. The predominance of branch and leaf shade in agricultural environments presents a barrier for accurate target recognition. Particularly for picking robots, precise localization of the picking object is essential. For this purpose, this paper proposes detection and localization methods based on deep learning and active sensing for harvesting robots in real-world environments with occlusion and varying lighting conditions. Using a deep learning network, the detection method firstly extracts the peduncle and fruit regions; the fruit region is then used to calculate the occlusion rate and the offset distance of the peduncle relative to the fruit. With such information, the robot arm adjusts the camera's field of view to perform multiple recognitions until the confidence is satisfied. Furthermore, to solve the picking problem caused by the peduncle's random tilting, this paper proposes a method to calculate the peduncle's tilt angle for controlling the end-effector to make the corresponding angle rotation. The robot arm and its end-effector are directed to complete the harvesting with the picking point location and tilt angle. In this study, data collection, detection and picking tests were implemented in the field, the results indicated that the method obtained an average successful picking rate of 90% after 300 trials, the error between the estimated occlusion ratio and the genuine value is 16% in average, and the active sensing method has improved the confidence score in occluded situations by over 50%. The proposed active methods have a 33% increase in precision and a 43% increase in efficiency compared to constant methods. [ABSTRACT FROM AUTHOR]

• A method for identifying corn seedlings and weeds based on hyperspectral images was proposed. • A spatial-spectral attention-enhanced Res-3D-OctConv model was proposed. • The proposed method has higher precision than traditional methods and deep learning methods. Corn production is an important basis to ensure the world food security, and weeds in the field will cause corn production decline. Therefore, in order to quickly recognize corn and weed in the field, a model was proposed by combining hyperspectral image with deep learning method. However, there are some problems in hyperspectral image, such as high redundancy of adjacent spectra and insufficient feature information extraction. In order to solve the above problems, the four principal components based on principal component analysis (PCA) were firstly extracted in this paper, so as to decrease the information redundancy between adjacent spectra. Secondly, the residual three-dimensional octave convolution (Res-3D-OctConv) was used to excavate the spatial information from the frequency components, while taking into account the spectral information. Finally, spatial and spectral attention models were introduced to highlight important spatial information and spectral information. At the same time, the spatial information and spectral information was integrated by cross fusion. Experimental results show that the recognition accuracy of the proposed model is 98.56 %, which is 8.65 % and 10.20 % higher than that of k-nearest neighbor (KNN) and support vector machine (SVM) respectively. The recognition result of the proposed model is further compared with that of 3D residual network (3D-ResNet) and 3D convolutional neural network (3D-CNN), and the recognition precision of the proposed model in this paper is increased by 1.40 % and 1.02 % compared with 3D-CNN and 3D-ResNet, respectively. The results show that the proposed model can better recognize the hyperspectral images of corn and weed. [ABSTRACT FROM AUTHOR]