The identification of plant nutritional stress based on visual symptoms is predominantly done manually and is performed by trained specialists to identify such anomalies. In addition, this process tends to be very time consuming, has a variability between crop areas and is often required for analysis at various points of the property. This work proposes an image recognition system that analyzes the nutritional status of the plant to help solve these problems. The methodology uses deep learning that automates the process of identifying and classifying nutritional stress of Brachiaria brizantha cv. marandu. An image recognition system was built and analyzes the nutritional status of the plant using the digital images of its leaves. The system identifies and classifies Nitrogen and Potassium deficiencies. Upon receiving the image of the pasture leaf, after a classification performed by a convolutional neural network (CNN), the system presents the result of the diagnosed nutritional status. Tests performed to identify the nutritional status of the leaves presented an accuracy of 96%. We are working to expand the data of the image database to obtain an increase in the accuracy levels, aiming at the training with a larger amount of information presented to CNN and, thus, obtaining results that are more expressive.

One possibility to perform nutritional assessment on plants is by analyzing the symptoms visually presented on their leaves. This evaluation is carried out by the technique of leaf diagnosis by specialized people, most of which is done manually, thus requiring specialized labor, which makes it difficult to use, especially in the Amazon region where this labor is still scarce. The objective of this study was the development and apply of Convolutional Neural Networks (CNN) models, which perform the classification of the Brachiaria brizantha cv. Marandu nutritional status using the image of its leaves. Six CNN models were implemented and evaluated: one based on AlexNet and the pre-trained VGG-16, VGG-19, Inception-V3, ResNet-50 and MobileNetV2.All pre-trained models used transfer of learning that saves time and obtains a better result in the identification of deficiencies. To classify them, a set of image data was created, both for deficient and healthy leaves, grown in a greenhouse to serve as the information to be learned by the models during training. These models classify the deficiencies of Potassium, Nitrogen and Phosphorus, in addition to identifying whether the plant is healthy. This technology can improve the production of Brazilian pastures and, consequently, improve the number of animals per pasture area, thus contributing to a more sustainable production, especially in the Amazon region. Of all the models tested, the best accuracy was VGG-16 with 96.93%, in test data.