Detecção de doenças foliares em imagens com Aprendizado Profundo: estudos de caso para míldio em soja e ácaro em lichieira

Abstract

The detection of foliar diseases is an important task, because the more efficient and effective this activity is, the greater the chances of success in the treatment of the plant. In this context, the development of techniques that automate this activity is relevant. This work aims to analyze the performance of Deep Learning techniques in the automatic classification of diseases in images of Lichieira and Soybean leaves. Soybean was chosen because it plays a fundamental role in the social and economic development of several regions of Brazil, which is one of the world’s largest producers of this grain. Lichieira was selected because it is a plant grown mainly in Brazil, being a tree for which there is a shortage of alternatives for the control and management of pests and diseases. The methodology used to carry out this work can be summarized in two main activities: i) production of the experimental database, and ii) Conducting a case study with Convolutional Neural Networks. The database produced is composed of 2288 images: 268 soybean leaves with mildew (a disease caused by the fungus Peronospora Manshurica), 1000 photos of lychee leaves with injuries caused by the Ehinosis mite (Aceria Litchii) and 1020 images of lychee leaves without the effect of this mite. It is noteworthy that the amount of soybean images is lower due to the low rainfall in the Dois Vizinhos region (at the end of 2021), which impacted productivity losses and the low proliferation of diseases that require moisture. After collecting these data, the images were labeled by a specialist with a doctorate in Phytopathology. With the literature review carried out, it was possible to identify that the PlantVillage database (which has 54305 images, divided into 38 classes) is the main base explored in works related to the detection of diseases in images with similar characteristics to those collected in this work – photos taken individually for each sheet with a uniform background. It was also possible to identify five Convolutional Neural Networks (AlexNet, DenseNet169, Inceptionv3, ResNet34 and VGG13) that present greater than 99.2% accuracy with the PlantVillage database. These five models were selected for the case study of this work, which was also conducted with three other more recent Convolutional Neural Networks: EfficientNetB7, MobilenetV2, and SqueezeNet. For each of the eight selected neural networks, experiments were carried out with three training strategies: i) fine-tuning (updating only the fully connected layers of a pre-trained CNN); ii) training from scratch (applies fine-tuning to all layers of the network and starts optimization back propagation from the pre-trained network); and iii) deep training (consists of training the network from scratch using a random configuration of weights, not using transfer learning). The experiments of this work were carried out in three stages, each with a different instance of the database. At all stages, 80% of the images made up the training set, and 20% of them were part of the test set. In the first stage, the eight RNCs were trained with a set of 57308 images: 2288 photos of Soybean and Lychee leaves collected for this work, 54305 from PlantVillage, and 715 photos with random backgrounds (which is a strategy used in related works to obtain more accurate accuracy in a deep neural network). For each of the training strategies applied in the first stage of the experiments, the RNC that obtained the best accuracy was selected, which was applied in subsequent case studies. In the second case study, the RNCs were trained with the 2020 images of Lychee leaves collected for this project. In the third case study, a set of 5358 images was used: 5090 healthy soybean leaves from PlantVillage, and 268 soybean leaves with mildew collected for this study. In the first stage, the best result was obtained with DenseNet169 (with the deep training strategy) which obtained an accuracy of 99.75%. For the classification of only Lychee leaves (with and without mite), the best accuracy was 94.80%, achieved with DenseNet169 (with the fine-tuning training strategy). As for the third case study, carried out only with soybean leaves, an accuracy of 100% was also obtained with DenseNet169 (in two training strategies: fine-tuning and deep training). With the results obtained, it was possible to analyze positive points and limitations of the case study and to propose themes for future work that can contribute to improving the results obtained.