Search

Water stress has been one of the most serious threat to the crop growth, development and yield quality in agricultural fields. Timely and accurate diagnosis of crop water stress can greatly contribute to the precision irrigation for the crop resilience and yield. In this study, the research object was taken from the summer maize in the typical dryland agricultural area of northwest China. A six-channel multispectral sensor was mounted on a drone to obtain the remote sensing image data of summer maize at the nodulatione and staminate pulling stage in 2022. At the same time, the stomatal conductance and phenotypic parameters of summer maize were also collected. The background was removed by supervised classification. The canopy vegetation index and image texture were obtained using the gray-scale covariance matrix. The sensitive vegetation index, image texture and phenotypic parameters and their combinations were screened out by the Bayesian information criterion and full subset filtering. The summer maize stomatal conductance estimation model was constructed to combine the three types of machine learnings: the extreme learning machine, the random forest, and the back-propagation neural network. The optimal model was mapped to estimate the stomatal conductance. The Pearson correlation coefficient of vegetation index and stomatal conductance were significantly positively correlated, whereas, the canopy reflectance of summer maize was weakly negatively correlated. Different types of image textures at different wavelengths were correlated with the stomatal conductance, and the highest correlation was found in the 550 nm band. The Pearson correlation coefficients between morphological structure phenotypes (plant height, stem thickness and leaf area) and stomatal conductance of summer maize were 0.72, 0.58 and 0.69, respectively, where the three types of phenotypic parameters data were correlated well with stomatal conductance. Vegetation indices with spectral reflectance data were used to assess the overall health and moisture status of the vegetation. Image texture was used to capture the spatial distribution, texture and structural features of crops. Crop phenotypic parameters were then used to reflect the physiological and morphological responses of the crop in a three-dimensional manner, providing visual information about the growth and moisture of the vegetation. The decision coefficients of the crop water stress diagnostic models that constructed from the three information sources increased from 0.728 and 0.750 to 0.841, respectively, compared with the single or two combinations, indicating the great potential to stomatal conductance prediction. The optimal combination of indicators was screened by Bayesian information criterion and full subset screening: DWSI, NDVI, MEA, ENT, plant height and leaf area. The back-propagation neural network model with the three complementary information sources was the optimal model for the water stress diagnosis of summer maize (coefficient of determination of 0.841, root mean square error of 0.043 mol/(m2 ·s), and mean absolute error of 0.034 mol/(m² ·s)). The underestimation of stomatal conductance was significantly improved, compared with the rest models. The inverse map with the optimal model was widely applied to easily and accurately diagnose the crop water stress for the purpose of irrigation strategies and resource allocation. The finding can provide a feasible and accurate diagnosis of water stress in summer maize. [ABSTRACT FROM AUTHOR]

Cotton is one of the most important economic crops in China. It is of great significance to predict the yield of cotton. In this study, the Unmanned Aerial Vehicle (UAV) remote sensing platform was first used to collect some data on the densely planted cotton in Xinjiang, China. Five-point sampling was also utilized in the period of defoliation. A total of five cotton fields were then selected as well. Secondly, each image was divided into multiple sub-images and then enhanced by color space conversion, brightness, noise blurring, flipping and rotating operation. A high-quality dataset was constructed to estimate the cotton yield. Better training was achieved, compared with the original. Eventually, the cotton dataset with the construction site was used to extract the element rate from the cotton images using RCH-UNet (resnet coordinate hardswish UNet). Among them, the UNet was used as a baseline model to construct an improved U-type convolutional neural network. While the IoU (intersection over union), PA (pixel accuracy), and precision were taken as the evaluation metrics. The overall RCH-UNet improved the three metrics by 14.34, 9.85, and 8.68 percentage points, respectively, compared with the original UNet. Specifically, ResNet50 backbone feature extraction network was selected to replace the traditional CBR (convolution batch normalization ReLU) downs sampling structure in UNet; The CA (coordinate attention) mechanism was combined with the UNet; The ReLU activation function in UNet was replaced with the hardswish activation function. The results showed that: 1) The richer semantic information in the ResNet50 was learned through the residual structure, indicating the better feature extraction and expression of the improved model. 2) The CA mechanism was significantly enhanced to learn the detailed features. At the same time, there was an effective reduction in the interference of irrelevant features to the model. 3) The hardswish activation function presented the stronger expression and feature fusion of UNet when performing up-sampling and jump connection. Subsequently, the performance of RCH-UNet was tested with the PSPNet and DeepLabv3 models under the same experimental conditions. The IoU, PA, and precision of RCH-UNet had improved by 9.15, 6.31, and 3.99 percentage points, respectively, compared with the DeepLabv3 model. A prediction model of cotton yield was constructed with the ridge regression from the cotton pixels extracted by the RCH-UNet model. The image texture features were also extracted by GLCM (gray-level co-occurrence matrix). The R² value of the improved model was 0.92, and the average relative error between the predicted and actual yield was 9.254%. The RCH-UNet model was accurately and effectively extracted from the cotton images; Meanwhile, the prediction model of cotton yield was effectively verified using deep learning and image processing. The UAV low-altitude imaging can also be expected to predict the yield of densely planted cotton in Xinjiang, China. [ABSTRACT FROM AUTHOR]

Ultra-high-resolution aerial images obtained from Unmanned Aerial Vehicles (UAVs) have widely been used to extract crop planting information in recent years. However, some high-resolution multispectral or hyperspectral images were usually costly and time-consuming for data processing. Therefore, it is very necessary to effectively use easily accessible and low-cost high-resolution RGB images, particularly to eliminate the common noises (e.g., shadows and bare land) for a better extraction accuracy of crop planting. In this study, a high-precision extraction method was proposed to obtain the maize planting information using 1.8 cm resolution UAV aerial orthophotos (i.e., RGB images). The experimental maize farm was located in Southeast Africa, where images were taken at noon during the maize growing season. The classification features were also selected from the aspects of the spectrum, color space, and image texture. Then, five types of classification were selected to extract maize planting information, including Bayes, K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF) algorithms). Firstly, an object-oriented interpreting platform (eCognition9.0 software) was selected to calculate the space transform of Hue, Saturation, and Intensity (HSI) color, eight types of RGB texture, and five vegetation indices, including the normalized green-red difference index, red-green ratio index, vegetation color verification index, visible-band difference vegetation index, and excess green vegetation index. Then, three types of feature space were constructed: 1) The first feature space was composed of three sub-feature spaces, i.e., vegetation indices, HSI color space features, and RGB image texture features; 2) The second feature space was composed of four sub-feature spaces, where three types of features were combined (i.e., vegetation indices, HSI color space, and RGB image texture) in pairs or total; 3) The third feature space was composed of the most optimal factors, where the dimension reduction was performed on the combination of all three types of features using RF. Subsequently, the RGB images were classified into three land-use types, including maize, bare land, and shadow. Bayes, KNN, SVM, DT, and RFs models were finally selected for the supervised classification with error matrix. The results showed that the optimal classification accuracy was obtained using neither a single feature nor all three types of features in total. More importantly, a combination of features was usually achieved higher accuracy than that of a single feature. Specifically, the best choice was the combination of HSI color and RGB image texture features using the RF, particularly with the total highest accuracy of 0.862 and a Kappa coefficient of 0.793. Additionally, the dimension reduction of features using RF models was neither significantly improved nor reduced classification accuracy (except for the SVM). However, the factors retained from the feature dimension reduction were easily explained suitable for the actual background and meaning. Furthermore, both classification efficiency and stability were improved greatly during this time. The finding can provide a specific solution for the high-precision extraction of crop planting information using UAV RGB images [ABSTRACT FROM AUTHOR]

With the continuous development of hyperspectral remote sensing technology, it has been widely used in vegetation mapping. However, sparse vegetation canopy, soil background, and spectral similarity between different types of vegetation are still the main challenges for vegetation types mapping in arid areas. As a result, it is difficult to achieve reliable classification accuracy by using spectral or texture features separately. Generalized Normal Distribution Optimization (GNDO) is a new feature optimization algorithm, with advantages in quality and stability of feature extraction results, comparing to traditional optimization algorithms. But it has not yet been applied to select bands of hyperspectral data. In order to validate the feasibility of combining ZY-1 02D spectral and texture features to classify vegetation types in arid areas, to verify the effectiveness of the GNDO method for bands selection of hyperspectral data, and to explore the effects of feature selection methods and training sample numbers on the classification accuracy of vegetation mapping, different Wrapper Optimization methods, such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), and GNDO, were applied to select spectral features for vegetation mapping, taking the area around Zongjia Town, Dulan County, Qinghai Province, China as the research area, then the band selection results of these methods were analyzed. Train sample set containing 30, 50, 100, 150, and 200 pixels per class were used to select bands and to train the classifier. Different methods (ALL (without bands selection), GA, PSO, GWO, GNDO) and different sizes of the trained sample sets (30, 50, 100, 150, and 200 pixels per class) were used to obtain 25 spectral feature-based classification data sets. Simultaneously, 8 texture features (mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment) were extracted using the Gray-level co-occurrence matrix (GLCM) method, and selected on basis of distinguishability for vegetation types. Texture features (TEX) were combined with spectral feature-based classification data sets. The random forest classification method was applied to classify vegetation types for the classification data sets, and the classification accuracy of classification data sets was evaluated and compared. The results show that 1) the blue region (400-450 nm), the red edge region (700-750 nm), and the red region (600-650 nm) are the most sensitive to distinguish the vegetation types in the study area; 2) the GNDO200 achieved the highest overall classification accuracy (80.44%) among the spectral feature-based classification data sets, which was better than the classification accuracy (78.86%) using all bands (ALL200); 3) with the increase of training samples, the overall classification accuracy of each classification data set showed an increasing trend, the classification accuracy of different feature selection methods showed different reliance on the number of training samples; 4) image texture features significantly improved the classification accuracy, and the GWO200+TEX dataset had the highest overall classification accuracy (82.86%). This study could verify the potential of the ZY1-02D, the new hyperspectral satellite data, for the classification of vegetation types in arid areas, and provide an idea for the selection of spectral and texture features in hyperspectral vegetation mapping. [ABSTRACT FROM AUTHOR]

A recirculating aquaculture system (RAS) has widely been acknowledged as eco-friendly land-based aquaculture since only a limited amount of water is used in recirculation. There is also much greater potential in modern intensive aquaculture, compared with conventional fish production. A key feature of RAS is the high precise controlling of the water environment in the rearing tank and the feeding of rearing fish. Water environmental parameters (such as water temperature, and oxygen levels) can provide stable and optimal conditions for less stress and better growth during fish living. In addition, fish feeding plays a vital role in RAS, because starvation may impact the organizational structures of fish, whereas, overfeeding may threaten the survival of animals, as well as the water quality. Therefore, it is essential to investigate the strategy of fish feeding in RAS. Currently, the feeding events are mainly implemented by the feeding machine with predetermined and fixed feeding times and quantities, according to the experience rather than the real fish appetite. Thus, it is necessary to evaluate the fish behavior pattern and appetite in the feeding strategy for a higher performance of RAS. In this study, an innovative evaluation of fish feeding behavior was proposed using Fourier spectrum feature extraction and support vector machines (SVM) classification. The degree of surface ripple dispersion was considered to predict the fish feeding behavior, rather than the only movement of fish bodies in conventional cases. Firstly, a high-speed video camera was used to capture the process before and after feeding. The video images were then converted from RGB to HSV space, where the background including the water ripple was subtracted, and the contrast was enhanced via the thresholds of saturation and brightness. Secondly, fast Fourier transform (FFT) was adopted to transform the image information from the spatial to a frequency domain. The fluctuation level of ripples in the image was regarded as an indicator of fish appetite using the amplitudes in the frequency spectrum. Specifically, the frequency spectrum was characterized by a high frequency, short period, and a large sum of amplitudes, when the fishes presented high appetite, and vice versa. Thirdly, the database of feature vectors was generated using the self-developed loop filters with various diameters. Finally, the obtained feature vector was trained to SVM. The feature vector of the SVM model was then optimized, where the optimal kernel width and penalty coefficient were selected to efficiently judge the feeding desire of fish during the cross validation. Moreover, a comparison was conducted to verify the efficiency, among the proposed, optical flow, image texture, shape, and texture features. As such, the original interference condition was transformed into the judgment, indicating the better universality of RAS against the high fish density, unstable water surface ripple, and dark light in the system. A field test also demonstrated that a high accuracy of 99.24% was achieved in judging the feeding behavior of farmed fish in RAS. In practice, it is strongly recommended to take videos of farmed fish as the training data for the classifier of fish feeding behavior. Consequently, the improved RAS can also be incorporated into the feeding machine to supervise the real-time smart and precise feeding events in modern aquaculture. [ABSTRACT FROM AUTHOR]