Artificial Intelligence in Agriculture最新文献

Artificial intelligence and machine learning have been increasingly applied for prediction in agricultural science. However, many models are typically black boxes, meaning we cannot explain what the models learned from the data and the reasons behind predictions. To address this issue, I introduce an emerging subdomain of artificial intelligence, explainable artificial intelligence (XAI), and associated toolkits, interpretable machine learning. This study demonstrates the usefulness of several methods by applying them to an openly available dataset. The dataset includes the no-tillage effect on crop yield relative to conventional tillage and soil, climate, and management variables. Data analysis discovered that no-tillage management can increase maize crop yield where yield in conventional tillage is <5000 kg/ha and the maximum temperature is higher than 32°. These methods are useful to answer (i) which variables are important for prediction in regression/classification, (ii) which variable interactions are important for prediction, (iii) how important variables and their interactions are associated with the response variable, (iv) what are the reasons underlying a predicted value for a certain instance, and (v) whether different machine learning algorithms offer the same answer to these questions. I argue that the goodness of model fit is overly evaluated with model performance measures in the current practice, while these questions are unanswered. XAI and interpretable machine learning can enhance trust and explainability in AI.

The increasing gap between the demand and productivity of maize crop is a point of concern for the food industry, and farmers. Its' susceptibility to diseases such as Turcicum Leaf Blight, and Rust is a major cause for reducing its production. Manual detection, and classification of these diseases, calculation of disease severity, and crop loss estimation is a time-consuming task. Also, it requires expertise in disease detection. Thus, there is a need to find an alternative for automatic disease detection, severity prediction, and crop loss estimation. The promising results of machine learning, and deep learning algorithms in pattern recognition, object detection, and data analysis motivate researchers to employ these techniques for disease detection, classification, and crop loss estimation in maize crop. The research works available in literature, have proven their potential in automatic disease detection using machine learning, and deep learning models. But, there is a lack none of these works a reliable and real-life labelled dataset for training these models. Also, none of the existing works focus on severity prediction, and crop loss estimation. The authors in this manuscript collect the real-life dataset labelled by plant pathologists. They propose a deep learning-based framework for pre-processing of dataset, automatic disease detection, severity prediction, and crop loss estimation. It uses the K-Means clustering algorithm for extracting the region of interest. Next, they employ the customized deep learning model ‘MaizeNet’ for disease detection, severity prediction, and crop loss estimation. The model reports the highest accuracy of 98.50%. Also, the authors perform the feature visualization using the Grad-CAM. Now, the proposed model is integrated with a web application to provide a user-friendly interface. The efficacy of the model in extracting the relevant features, a smaller number of parameters, low training time, high accuracy favors its importance as an assisting tool for plant pathology experts.The copyright for the associated web application ‘Maize-Disease-Detector’ is filed with diary number: 17006/2021-CO/SW.

Weed identification is fundamental toward developing a deep learning-based weed control system. Deep learning algorithms assist to build a weed detection model by using weed and crop images. The dynamic environmental conditions such as ambient lighting, moving cameras, or varying image backgrounds could affect the performance of deep learning algorithms. There are limited studies on how the different image backgrounds would impact the deep learning algorithms for weed identification. The objective of this research was to test deep learning weed identification model performance in images with potting mix (non-uniform) and black pebbled (uniform) backgrounds interchangeably. The weed and crop images were acquired by four canon digital cameras in the greenhouse with both uniform and non-uniform background conditions. A Convolutional Neural Network (CNN), Visual Group Geometry (VGG16), and Residual Network (ResNet50) deep learning architectures were used to build weed classification models. The model built from uniform background images was tested on images with a non-uniform background, as well as model built from non-uniform background images was tested on images with uniform background. Results showed that the VGG16 and ResNet50 models built from non-uniform background images were evaluated on the uniform background, achieving models' performance with an average f1-score of 82.75% and 75%, respectively. Conversely, the VGG16 and ResNet50 models built from uniform background images were evaluated on the non-uniform background images, achieving models' performance with an average f1-score of 77.5% and 68.4% respectively. Both the VGG16 and ResNet50 models' performances were improved with average f1-score values between 92% and 99% when both uniform and non-uniform background images were used to build the model. It appears that the model performances are reduced when they are tested with images that have different object background than the ones used for building the model.

Artificial intelligence (AI) has advanced at an astounding rate and transformed numerous economic sectors. Nevertheless, a comprehensive understanding of how AI can improve the agri-food industry is lacking. In addition, there is a notable dearth of research on AI that investigates the influence of AI on agri-food resources and educates practitioners on the significance of knowledge-based and smart agriculture. We utilised bibliometric analysis to investigate the present state of the art and emerging trends in the relationship between AI and the agri-food industry. The research identified three distinct growth phases and the most prevalent AI strategies in the industry. In addition, we analysed key trends and offered researchers and practitioners insightful recommendations for future research. Using resource-based view (RBV) as the theoretical lens, this study established a framework emphasising the long-term effects of AI on various agri-food resources and proposed several research propositions. In addition, AI-related obstacles have been identified and categorised into four major categories. Lastly, the originality of the article lies in its numerous research suggestions and recommendations for advancing the AI field in the agri-food industry.

Plants diseases have a detrimental effect on the quality but also on the quantity of agricultural production. However, the prediction of these diseases is proving the effect on crop quality and on reducing the risk of production losses. Indeed, the detection of plant diseases -either with a naked eye or using traditional methods- is largely a cumbersome process in terms of time, availability and results with a high-risk error. The present work introduces a depth study of various CNN architectures with different optimization algorithms carried out for olive disease detection using classification techniques that recommend the best model for constructing an effective disease detector. This study presents a dataset of 5571 olive leaf images collected manually on real conditions from different regions of Morocco, that also includes healthy class to detect olive diseases. Further, one of the goals of this research was to study the correlation effects between CNN architectures and optimization algorithms evaluated by the accuracy and other performance metrics. The highest rate in trained models was 100 %, while the highest rate in experiments without data augmentation was 92,59 %. Another subject of this study is the influence of the optimization algorithms on neuronal network performance. As a result of the experiments carried out, the MobileNet architecture using Rmsprop algorithms outperformed the others combinations in terms of performance and efficiency of disease detector.

In the last few years, deep neural networks have achieved promising results in several fields. However, one of the main limitations of these methods is the need for large-scale datasets to properly generalize. Few-shot learning methods emerged as an attempt to solve this shortcoming. Among the few-shot learning methods, there is a class of methods known as embedding learning or metric learning. These methods tackle the classification problem by learning to compare, needing fewer training data. One of the main problems in plant diseases and pests recognition is the lack of large public datasets available. Due to this difficulty, the field emerges as an intriguing application to evaluate the few-shot learning methods. The field is also relevant due to the social and economic importance of agriculture in several countries. In this work, datasets consisting of biotic stresses in coffee leaves are used as a case study to evaluate the performance of few-shot learning in classification and severity estimation tasks. We achieved competitive results compared with the ones reported in the literature in the classification task, with accuracy values close to 96%. Furthermore, we achieved superior results in the severity estimation task, obtaining 6.74% greater accuracy than the baseline.

Performing accurate and automated semantic segmentation of vegetation is a first algorithmic step towards more complex models that can extract accurate biological information on crop health, weed presence and phenological state, among others. Traditionally, models based on normalized difference vegetation index (NDVI), near infrared channel (NIR) or RGB have been a good indicator of vegetation presence. However, these methods are not suitable for accurately segmenting vegetation showing damage, which precludes their use for downstream phenotyping algorithms. In this paper, we propose a comprehensive method for robust vegetation segmentation in RGB images that can cope with damaged vegetation. The method consists of a first regression convolutional neural network to estimate a virtual NIR channel from an RGB image. Second, we compute two newly proposed vegetation indices from this estimated virtual NIR: the infrared-dark channel subtraction (IDCS) and infrared-dark channel ratio (IDCR) indices. Finally, both the RGB image and the estimated indices are fed into a semantic segmentation deep convolutional neural network to train a model to segment vegetation regardless of damage or condition. The model was tested on 84 plots containing thirteen vegetation species showing different degrees of damage and acquired over 28 days. The results show that the best segmentation is obtained when the input image is augmented with the proposed virtual NIR channel (F1=$0.94$) and with the proposed IDCR and IDCS vegetation indices (F1=$0.95$) derived from the estimated NIR channel, while the use of only the image or RGB indices lead to inferior performance (RGB(F1=$0.90$) NIR(F1=$0.82$) or NDVI(F1=$0.89$) channel). The proposed method provides an end-to-end land cover map segmentation method directly from simple RGB images and has been successfully validated in real field conditions.

A reliable and accurate forecasting model for crop yields is crucial for effective decision-making in every agricultural sector. Machine learning approaches allow for building such predictive models, but the quality of predictions decreases if data is scarce. In this work, we proposed data-augmentation for wheat yield forecasting in the presence of small data sets of two distinct Provinces in Algeria. We first increased the dimension of each data set by adding more features, and then we augmented the size of the data by merging the two data sets. To assess the effectiveness of data-augmentation approaches, we conducted three sets of experiments based on three data sets: the primary data sets, data sets with additional features and the augmented data sets obtained by merging, using five regression models (Support Vector Regression, Random Forest, Extreme Learning Machine, Artificial Neural Network, Deep Neural Network). To evaluate the models, we used cross-validation; the results showed an overall increase in performance with the augmented data. DNN outperformed the other models for the first Province with a Root Mean Square Error (RMSE) of 0.04 q/ha and R_Squared (R²) of 0.96, whereas the Random Forest outperformed the other models for the second Province with RMSE of 0.05 q/ha. The data-augmentation approach proposed in this study showed encouraging results.