查看更多>>摘要：? 2022 Elsevier B.V.There are few publicly available large-scale datasets for tomato leaf disease detection using deep learning, the workload associated with manual data collection is high, and conventional supervised data augmentation methods (such as image rotation, flipping, and shifting) are challenging to attain satisfactory results with. While Deep Convolutional Generative Adversarial Networks (DCGAN) is a popular unsupervised data augmentation method based on deep learning, the data augmentation method based on DCGAN may have several drawbacks, including poor image quality and excessive computing resource consumption. Given this, a method based on Fast WDBlock based GAN (FWDGAN) was proposed in this paper. For the network's generator, a wide and deep feature extraction block (WDBlock) with a two-path strategy was designed, combining the extracted depth feature based on ResNet and the extracted global feature based on InceptionV1. By incorporating WDBlock into the generator, the quality of the generated tomato leaf disease images was improved. For the network's discriminator, the Depthwise separable convolution Discriminator (DSC-Discriminator) that significantly reduced the model's parameters without impairing the network's performance was constructed. Finally, the SeLU activation function was used selectively to improve the training stability of the network. Comparative experiments demonstrated that FWDGAN could generate higher-quality data, with FID scores of 193.998, 264.704, 260.594, and 161.436 for healthy tomato leaf image, Leaf Mold, Septoria leaf spot, and Yellow Leaf Curl Virus generated by FWDGAN, respectively. Furthermore, the total number of parameters in FWDGAN was approximately one-third less than that of DCGAN.
查看更多>>摘要：? 2022 Elsevier B.V.A hydrostatic power split continuously variable transmission (CVT) for a tractor often has multiple ranges to improve transmission efficiency. However, a failure of the wet-clutch control system will lead to interruption of tractor power and even endanger driving safety. To improve the reliability of CVT tractors, methods for diagnosing faults in the wet-clutch control system were studied. A test bench was built and the clutch engagement pressure was measured under different fault modes. For these time series data, statistics such as the mean, variance, kurtosis etc. were selected as features in the analysis. An improved Gaussian naive Bayes algorithm based on a time window as well as principal component analysis were used to classify the different fault modes of the clutch control system. Finally, the performance of the algorithm was analyzed. A test with multiple faults was run, as was a comparison with traditional algorithms. By optimizing the time window for data interception, the classification accuracy of the Gaussian naive Bayes algorithm for normal operation reached 97% and reached 100% for the fault modes. The average accuracy and recall rate of fault diagnosis were 98.2% and 89.4%, respectively, which are better than the results for a support vector machine, the k-nearest neighbors algorithm, or the decision tree algorithm. Importantly, the results show that the clutch pressure fluctuations during the shift by a tractor CVT can be used for the fault diagnosis of a clutch control system.
查看更多>>摘要：? 2022 Elsevier B.V.Analytical grade ’in vivo’ plant metabolic quantification using spectroscopy is a key enabling technology for precision agriculture.Current methods such as PLS, ANN and LS-SVM are non-optimal for resolving spectral interference and matrix effects to provide similar results to the analytical chemistry laboratory. This research presents a new self-learning artificial intelligence (SL-AI) method based on the search of covariance modes. These isolate the different modes of interference present in spectral data, allowing the consistent quantification of constituents. A review of the state-of-the-art methods with the figures of merit mean absolute standard error percentage (MASEP) and Pearson correlation coefficient (R) is presented for comparison and discussion. 707 grapes were quantified for glucose, fructose, malic and tartaric acids in five wine-making and one table grape varieties, and used to benchmark the new method against the state-of-the-art methodologies: partial least squares, local partial least squares, artificial neural networks and least squares support vector machines. SL-AI provides consistent quantifications, whereas previous methods exhibit data-driven performance dependence. Pearson correlations of 0.93 to 0.99 and MASEP of 3.70% to 7.33% were obtained with the new methodology. Local partial least squares, the method with the best benchmarks from literature, achieved correlations of 0.81 to 0.94 and MASEP of 8.00% to 13.4%. The covariance mode isolates a particular interference, providing a direct relationship between spectral inference and constituent concentrations, consistent with the Beer–Lambert law. Such quantifies non-dominant absorbance constituents (e.g. sugars and acids), which is a significant step towards ’in vivo’ plant physiology-based precision agriculture.
查看更多>>摘要：? 2022 Elsevier B.V.Sentinel-2 satellite imagery offers a wealth of spectral information combined with a weekly temporal resolution. It is seen as a promising tool to extract spatial information about vineyards and link them to agronomic parameters. Usually, only one or a few images are commonly employed at specific stages like veraison in viticulture. Extracting further information from time-series images may be of interest; however, this remains an issue due to the noisy and complex nature of extracted time-series. The functional analysis proposes a robust continuous representation of these time-series, which can then be used with adapted statistical tools. This paper focuses on extracting relevant information at the within-field level on two vineyards in Spain, which can be jointly interpreted with field observations and measurements. More precisely, it discusses the use of popular linear dimensionality reduction techniques, namely Principal Component Analysis (PCA) and Partial Least Square (PLS), adapted to functional data in order to decompose NDVI time-series into a weighted sum of several functional components. The unsupervised methods, like PCA, decomposed the spatial structure within the vineyards using a few components, resulting in a better and more manageable dataset than the one obtained using simple non-constrained methods. The results show significant correlations with ground-truth data showing the added value of considering the whole NDVI temporal series compared to a single NDVI map at veraison. The proposed approach provided helpful information about each component's yearly trend. Moreover, the results are linked to grapevines' seasonal phenology and management practices, highlighting phenomena affecting the vineyard's development. This method is particularly suited for interactions with field experts, who may derive relevant agronomic information from the decomposition maps.
查看更多>>摘要：? 2022 The AuthorsHundreds of crop variety trial sites are operated across Australia with up to 100 small plots (20 m2) that require manual, labour-intensive monitoring for emergence, height, canopy cover and flowering status. Machine vision systems can reduce labour in monitoring, and infield fixed cameras are most suitable to provide low-cost continuous sensing without requiring labour of travelling to sites. Height is commonly detected using stereo cameras; however, a low-cost single camera system is preferable for broad scale use at variety trial sites. Existing low-cost fixed camera systems assess multiple plots in the image's field of view from a fixed mask that needs to be manually updated as the crop grows. Perspective transformation could be applied automatically to identify plot locations in the image. Existing systems also analyse multiple images independently and filter results to remove impacts from lighting variations. An alternative approach is to only analyse images in the same lighting conditions each day to reduce the need for filtering. In addition, a series of daily images can be used to track multiple leaf positions to monitor plant growth. A machine vision system has been developed that combines these technologies to track plant height from a series of daily images selected in the same lighting conditions, in combination with perspective transformation, in multiple plots in the camera's field of view. Emergence, canopy cover and flowering status were also identified at the canopy surface in each plot using colour segmentation and/or shape analysis. The algorithms were evaluated on 5 and 10 m towers monitoring randomised trials of 16 maize and soybean plots and detected maize flowering date within one day, soybean height (RMSE = 18.38 cm; R2 = 0.880), maize height (RMSE = 47.73 cm; R2 = 0.838), soybean canopy cover (RMSE = 22.14%; R2 = 0.818) and maize canopy cover (RMSE = 14.01%; R2 = 0.750). The larger error in maize height detection was due to the flowers being tracked by the algorithm instead of vegetation. Further work is required to transfer the algorithms to other crops and varieties.
查看更多>>摘要：? 2022 Elsevier B.V.The identification of individual dairy cows is an important prerequisite for dairy cow behaviour analysis and disease detection. Computer vision-based cow recognition is a noncontact and stress-free approach. In a free environment in a barn, due to changes in camera position and angle, recorded cow patterns are often deformed, making individual cow identification difficult. For cows in an unconstrained barn environment, this paper proposes a method for individual cow identification. First, a top-view image of a cow is obtained, and an improved Mask R-CNN is used to segment this image and extract the shape features of the cow's back. Then, a Fisher approach is used to select the best feature subset, and a support vector machine (SVM) classifier is applied to identify individual cows. To verify the effectiveness of the target detection algorithm, the proposed method is compared with the traditional Mask R-CNN model, and the precision, recall, F1 score, average run time per image and average precision of the improved Mask R-CNN model are 98.21%, 96.48%, 97.34%, 1.02 s, and 97.39%, respectively. An SVM classifier trained based on the obtained shape features is used for individual cow identification. The proposed method achieves a 98.67% cow identification accuracy based on a dataset containing top-view images of 48 cows. The results demonstrate the effectiveness of the proposed cow identification method and its significant potential for use in precision dairy cow management.
查看更多>>摘要：? 2022 Elsevier B.V.Rice planthoppers (RPH) are important pest that cause severe yield losses of rice production in China. South and Southwest China are the main infestation areas on the annual spread of RPH. Identifying the spatial and temporal patterns in migration and subsequent development can provide insight into underlying mechanisms driving the spread of RPH and assist governments with prioritizing areas to achieve proactive management and prevention. Across rice production regions in South and Southwest China, RPH population were recorded by light traps and field surveys at 195 counties from March to October in 2000 to 2019. We first measured the spatial patterns in RPH immigrant populations and field populations with spatial autocorrelation analysis. Then, spatial hotspot analysis was undertaken to highlight dense RPH regions, and significant hot spots were further extracted for comparing the spatial and temporal patterns according to RPH species. The results revealed that RPH population were highly aggregated in space over large geographical distances up to 700 km approximately. Geographic patterns and hot spots of populations varied substantially with time and species, and various spatial patterns might be determined by inherent properties of RPH and rice eco-systematic. Overall, these results provided essential information to improving and optimizing the monitoring network for RPH. The findings from this research may provide helpful information to enhance proactive management against RPH and offer sustainable management practitioners new opportunities to design, develop, and implement optimal pest control strategies to protect rice production in China.
查看更多>>摘要：? 2022 Elsevier B.V.Apple bruises, which are mainly caused by external forces during picking, transportation and marketing, could cause significant economic losses in postharvest storage. Considering the short time window (1–2 h) between bruises occurring and apples are stored, there needs a technique that can identify the invisible early stage bruises (defined as within 1–2 h after occurring), which cannot be completed by the conventional uniform imaging techniques. This study proposed and validated a new means, i.e., optical property mapping, for detecting apple bruises at the early stage. SFDI system was firstly validated, with the results showing that accuracy for measuring μa and μs' were within 8.71 % and 4.93 %, respectively. Optical property mappings of non-bruised and early stage bruised apples were extracted by the methods of three-phase coupled with curve fitting (TP-CF), and single snapshot coupled with look-up table (SS-LUT). The extracted μs' mappings were capable of detecting early stage bruises, and the TP-CF approach resulted into better performance than SS-LUT, but at the expense of lowering efficiency about 14,000 times. Denoising algorithms significantly enhanced the visual effect of bruise detection by minimizing the influence of apple peel spots. The approach proposed in this study is promising for detecting early stage bruising for apple industry.
查看更多>>摘要：? 2022The improvement of economic management in farms has become an important research topic in recent decades as the most dominant feature of current farm management information systems (FMIS). Production cost statistics allow farmers to assess the economic impact of farm activities and compare historical data against previous farm practices or competitors’ activities. Therefore, the availability of reliable cost data is of utmost importance for FMIS, especially data on agricultural machinery usage. Technical sheets, grey literature, and international standards provide estimates of farm operation costs, but they suffer from low accuracy because agricultural machinery is subjected to the high variability of both environmental and working conditions. Based on these considerations, this work aims to develop a novel methodology for cost calculations of field operations harnessing real-world CANBUS data based on the activity-based costing (ABC) approach. The research was conducted on a 198-kW tractor equipped with a CANBUS logger and several implements on which Bluetooth beacons were installed to automatically recognise agricultural operations. The acquired data were processed to identify the daily jobs performed by observing machine position (e.g., field, farm, or road) and operating condition states (e.g., moving, fieldwork, or idling). The ABC approach was applied in two steps: first, cost driver rates were assessed to define capital and non-capital costs; then, the costs of each agricultural operation performed were defined, correlating the cost drivers with the recorded jobs. The results show that fuel and labour costs combined affect 63%–71% of the total cost per hectare for the tested implements. The cost per hectare was found to be highly variable: the biggest gap between the higher and lower values registered with the same implement was 216.48 € ha?1. This methodology could help farmers to make more thoughtful decisions about crop, land, and farm operations management.
查看更多>>摘要：? 2022 Elsevier B.V.Citrus Huanglongbing (HLB) has posed a great challenge to the citrus production. Timely removal of HLB infected trees was considered as one of the most effective strategies for citrus orchard management. Therefore, rapid detection of HLB disease is urgently needed. The study was aimed to propose an effective method for HLB disease detection by developing a handheld device to capture multicolor fluorescence and multispectral reflectance images synchronously. Additionally, the deep learning and transfer learning technologies were introduced for citrus HLB disease detection. The results demonstrated that the lightweight convolutional neural network (MobileNetV3) can achieve an overall accuracy of 92.1% with the false negative rate of 12.1% at epochs of 33 by combining multicolor fluorescence with multispectral reflectance images as the input of MobileNetV3 model using the dataset of Navel orange. It implied that structural and physiological information from reflectance images and multicolor fluorescence images relating to photosynthesis and secondary metabolites were valuable for rapid HLB disease detection. The transfer learning method of fine-tuning model obtained a superior transferring ability than that of reuse-model with the overall accuracy of 96.5% for Ponkan. These results demonstrated the feasibility of developed handheld device based on multicolor fluorescence and multispectral reflectance imaging combined with deep learning and transfer learning technologies for high throughput HLB disease detection in different infected statuses and cultivars.
NSTL
Elsevier