Detecção e classificação de doenças e pragas da soja usando imagens de veículos aéreos não tripulados e técnicas de visão computacional

Abstract

Soybean (Glycine max) is an oleaginosa with good nutritional profile and important economic support for the food industry around the world. From sowing to harvesting, soybean cultivation can be affected by biotic pest and disease attacks, which may lead to depreciation of the final product. This thesis proposes to create a new system of computational vision to identify diseases and pests in the soybean crop that is able to present to the producer the level of infestation of each region of the field for a more efficient management of its populations. The proposed system is based on the Simple Linear Iterative Clustering (SLIC) method to segment plant leaves in the planting images captured by Unmanned Aerial Vehicles (UAVs) from a very popular UAV market line, called DJI Phantom. Then, shallow and deep techniques of machine learning are used in the classification of diseases and agricultural pests. In the task of disease classification, the performance of supervised machine learning techniques was evaluated for five different flight heights: 1, 2, 4, 8 and 16 meters. With shallow techniques (SVM, k-NN, Naive Bayes, J48, AdaBoost and Random Forest) and visual attributes of the leaves (color, gradient, texture and shape), the experiments supported the hypothesis that the closest flight heights between the UAV and the plant show higher correct classification rates (CCR). For shallow learning, specifically using the SVM classifier, the experiments reached 98.34% in the identification of foliar disease using flight heights between 1 and 2 meters, with decay of 2% to each meter, since the resolution of the plant in the image decreases to a higher flight height. The experiments also confirmed that even cataloging new diseases and increasing the number of 3 to 6 classes, there was no evidence of significant statistical difference of the performance metrics for the data collected by UAV at 2 meters height of the plantation. With deep learning techniques, four models recognized by the competition in ImageNet were evaluated for the automatic recognition of soybean leaf diseases using different strategies of fine-tuning and transfer learning on a set of 3,000 images captured under real field conditions. Data augmentation and dropout were used during network training to increase the amount of data and randomly eliminate some of the hidden neurons in the network, avoiding overfitting. The results showed that the Inception-V3, Resnet-50, VGG-19 and Xception models achieved higher classification rates using the fine-tuning strategy, achieving 99.04% of correctly classified samples. Likewise, the deep learning models were evaluated for the classification task of soybean insects using different strategies of fine-tuning and transfer learning on a set of 5,000 images collected directly from the field with the camera of a smartphone. The results showed that deep learning models trained with a fine-tuning achieved the highest classification rates compared to other approaches, reaching an accuracy of up to 93.82%. In addition, deep learning models have outperformed traditional features extraction methods such as SIFT and SURF with the bag of visual words approach, the OPFSEMImst semi-supervised learning method, and the supervised learning methods used to classify images, such as SVM, k-NN, Naive Bayes, J48, AdaBoost and Random Forest. In the pest insect counting task, three deep learning models were evaluated. The weights of the DenseNet-201 FT convectional neural network (CNN) - which obtained the highest value for accuracy (94.89%) in the classification task - were used in our computer vision system for counting insects in the image obtained by adding the superpixels of each class. In this way, we demonstrate how a deep learning model can be implemented in an end-to-end computer vision system to operate in a real field setting, under different lighting conditions, object size and background variations. In the post-processing stage a color map is generated, providing a class of each problem (disease or pest) per segment and the sum of the superpixels presents to the producer the level of infestation of a crop region by classifying the segments of an image of the plantation. The results indicate that the deep learning models can be successfully used to support specialists and farmers in the monitoring of pathogens and pests in soybean fields.