Search

A key challenge for autonomous precision weeding is to reliably and accurately detect weed plants and crop plants in real time to minimise damage to surrounding crop plants while performing weeding actions. Specifically for a wheat farm, classifying ryegrass weed plants is particularly difficult even with human eyes since ryegrass shows visually very similar shape and texture to the crop plants themselves. A Deep Neural Network (DNN) that exploits the geometric location of ryegrass is proposed for the real time segmentation of inter-row ryegrass weeds in a wheat field. Our proposed method introduces two subnets in a conventional encoder-decoder style DNN to improve segmentation accuracy. The two subnets treat inter-row and intra-row pixels differently, and provide corrections to preliminary segmentation results of the conventional encoder-decoder DNN. A dataset captured in a wheat farm by an agricultural robot at different time instances is used to evaluate the segmentation performance, and the proposed method performs the best among various popular semantic segmentation algorithms. The proposed method runs at 48.95 Frames Per Second (FPS) with a consumer level graphics processing unit, thus is real-time deployable at camera frame rate.

HighlightsBiGRU-attention based cow behavior classification was proposed.Key spatial-temporal features were captured for behavior representation.BiGRU-attention achieved >82% classification accuracy on calf and adult cow datasets.The proposed method could be used for similar animal behavior classification.Abstract. Animal behavior consists of time series activities, which can reflect animals’ health and welfare status. Monitoring and classifying animal behavior facilitates management decisions to optimize animal performance, welfare, and environmental outcomes. In recent years, deep learning methods have been applied to monitor animal behavior worldwide. To achieve high behavior classification accuracy, a BiGRU-attention based method is proposed in this article to classify some common behaviors, such as exploring, feeding, grooming, standing, and walking. In our work, (1) Inception-V3 was first applied to extract convolutional neural network (CNN) features for each image frame in videos, (2) bidirectional gated recurrent unit (BiGRU) was used to further extract spatial-temporal features, (3) an attention mechanism was deployed to allocate weights to each of the extracted spatial-temporal features according to feature similarity, and (4) the weighted spatial-temporal features were fed to a Softmax layer for behavior classification. Experiments were conducted on two datasets (i.e., calf and adult cow), and the proposed method achieved 82.35% and 82.26% classification accuracy on the calf and adult cow datasets, respectively. In addition, in comparison with other methods, the proposed BiGRU-attention method outperformed long short-term memory (LSTM), bidirectional LSTM (BiLSTM), and BiGRU. Overall, the proposed BiGRU-attention method can capture key spatial-temporal features to significantly improve animal behavior classification, which is favorable for automatic behavior classification in precision livestock farming. Keywords: BiGRU, Cow behavior, Deep learning, LSTM, Precision livestock farming.

Achieving rapid and accurate detection of apple leaf diseases in the natural environment is essential for the growth of apple plants and the development of the apple industry. In recent years, deep learning has been widely studied and applied to apple leaf disease detection. However, existing networks have too many parameters to be easily deployed or lack research on leaf diseases in complex backgrounds to effectively use in real agricultural environments. This study proposes a novel deep learning network, YOLOX-ASSANano, which is an improved lightweight real-time model for apple leaf disease detection based on YOLOX-Nano. We improved the YOLOX-Nano backbone using a designed asymmetric ShuffleBlock, a CSP-SA module, and blueprint-separable convolution (BSConv), which significantly enhance feature-extraction capability and boost detection performance. In addition, we construct a multi-scene apple leaf disease dataset (MSALDD) for experiments. The experimental results show that the YOLOX-ASSANano model with only 0.83 MB parameters achieves 91.08% mAP on MSALDD and 58.85% mAP on the public dataset PlantDoc with a speed of 122 FPS. This study indicates that the YOLOX-ASSANano provides a feasible solution for the real-time diagnosis of apple leaf diseases in natural scenes, and could be helpful for the detection of other plant diseases.

Knowing where livestock are located enables optimized management and mustering. However, Australian farms are large meaning that many of Australia's livestock are unmonitored which impacts farm profit, animal welfare and the environment. Effective animal localisation and counting by analysing satellite imagery overcomes this management hurdle however, high resolution satellite imagery is expensive. Thus, to minimise cost the lowest spatial resolution data that enables accurate livestock detection should be selected. In our work, we determine the association between object detector performance and spatial degradation for cattle, sheep and dogs. Accurate ground truth was established using high resolution drone images which were then downsampled to various ground sample distances (GSDs). Both circular and cassegrain aperture optics were simulated to generate point spread functions (PSFs) corresponding to various optical qualities. By simulating the PSF, rather than approximating it as a Gaussian, the images were accurately degraded to match the spatial resolution and blurring structure of satellite imagery. Two existing datasets were combined and used to train and test a YoloV5 object detection network. Detector performance was found to drop steeply around a GSD of 0.5m/px and was associated with PSF matrix structure within this GSD region. Detector mAP performance fell by 52 percent when a cassegrain, rather than circular, aperture was used at a 0.5m/px GSD. Overall blurring magnitude also had a small impact when matched to GSD, as did the internal network resolution. Our results here inform the selection of remote sensing data requirements for animal detection tasks, allowing farmers and ecologists to use more accessible medium resolution imagery with confidence., Comment: 20 pages, 9 figures, 4 tables in appendix

Deep learning methods such as convolutional neural networks (CNN) have become popular for addressing crops and weeds classification problems in agricultural robotics. However, to have satisfactory performance and avoid overfitting, training deep neural nets typically requires thousands of labeled images. This leads to tedious pixelwise labeling for semantic segmentation. In this paper, we hinge on the recent development in data augmentation and utilize the concept further for semantic segmentation and classification of crops and weeds. To be specific, we propose a novel data augmentation framework, based on the random image cropping and patching (RICAP) method, which is originally designed to augment data for generic image classification. The proposed framework introduces novel enhancements to the original RICAP so that it can be effectively used for data augmentation of semantic segmentation tasks. We evaluate the proposed methodology on two datasets from different farms. Comprehensive experimental evaluations and ablation studies show that the proposed framework can effectively improve segmentation accuracies, and the enhancements made over the original RICAP actually contribute to the performance gain. On average, the proposed method increases the mean accuracy and mean intersection over union (IOU) of the deep neural net with the conventional data augmentation (random flipping, rotation and colour jitter) from 91.01 to 94.02 and from 63.59 to 70.77 respectively for Narrabri dataset, and from 97.99 to 98.51 and from 74.26 to 77.09 respectively for Bonn dataset. The limitation of the proposed method, especially when a large number of training data is available, has also been discussed.

There has been an increasing demand for animal protein due to several factors such as global population growth, rising incomes, etc. However, farming productivity is stagnating due to a mix of traditional practice, climate change, socio-economic, and environmental phenomena. Precision livestock farming, with intelligent perception tools at its core, and vast amounts of data being acquired from different sensors or platforms, has the ability to analyse individual animal for improved management, and the potential to dramatically enhance farm productivity. In order to facilitate research and promote the development of related areas, this review summarises and analyses the main existing techniques used in precision cattle farming, focusing on those related to identification, body condition score evaluation, and live weight estimation. More than 100 relevant papers have been discussed in a cohesive manner. From this review and extensive discussions of recent trends, we anticipate that intelligent perception for precision cattle farming will develop through non-contact, high precision, automated technologies, combined with emerging 3D model reconstruction and deep learning technologies. Existing challenges and future research opportunities will also be highlighted and discussed.

Weeding is an effective way to increase crop yields. Reliable and accurate weed recognition is a prerequisite for achieving high-precision site-specific weed control in precision agriculture. To improve weed and crop recognition accuracy, a CNN feature based graph convolutional network (GCN) based approach is proposed. A GCN graph was constructed based on extracted weed CNN features and their Euclidean distances. Based on the semi-supervised learning, the GCN graph enriched the model by exploiting labeled and unlabeled image features, and testing samples obtain label information from labeled weed data by performing propagation over the graph. The proposed GCN-ResNet-101 approach achieved 97.80%, 99.37%, 98.93% and 96.51% recognition accuracies on four different weed datasets respectively, which outperformed the state-of-the-art methods (AlexNet, VGG16 and ResNet-101). Additionally, the runtime of the proposed approach also satisfies the real-time requirement of field weed control. The proposed CNN feature based GCN approach is favorable for multi-class crops and weeds recognition with limited labeled data, which is a promising approach in dealing with similar agricultural recognition tasks. Furthermore, the used datasets and source code are publicly available to facilitate the research in the recognition of field weeds.

In precision livestock farming, computer vision based approaches have been widely used to obtain individual cattle health and welfare information such as body condition score, live weight, activity behaviours. For this, precisely segmenting each cattle image from its background is a prerequisite, which is an important step towards obtaining real-time individual cattle information. In this paper, an instance segmentation approach based on a Mask R-CNN deep learning framework is proposed to solve cattle instance segmentation and contour extraction problems in a real feedlot environment. The proposed approach consists of the following steps: key frame extraction (detect the huge cattle motion frames), image enhancement (reduce the illumination and shadow influence), cattle segmentation and body contour extraction. We trained and tested the proposed approach on a challenging cattle image dataset. According to the experimental results, the proposed approach can render fairly desirable cattle segmentation performance with 0.92 Mean Pixel Accuracy (MPA) and achieve contour extraction with an Average Distance Error (ADE) of 33.56 pixel, which is better than that of the state-of-the-art SharpMask and DeepMask instance segmentation methods.