Avaliação do few shot learning para classificação de imagens de produtividade da soja obtidas por aeronave remotamente pilotada

Abstract

Given the growing importance of soy in the global agricultural economy and the need to improve agricultural practices to maintain the balance between food security and

the environment, this study sought to evaluate the effectiveness of using a sub- concept of Deep Learning (DL) the method of Few Shot Learning (FSL), applied at

the time to classify productivity images of soybean cultivation, acquired using a remotely piloted aircraft. RGB images were used in two different resolutions, 10 cm/px and 26 cm/px, obtained on the same day. After pre-processing the images, including class balancing, a database was obtained containing 9,721 images distributed into four soybean productivity classes: low, medium, high and very high. The following algorithms were explored: modified generic Convolutional Neural Network (CNN), resNet50 and denseNet121 applying the concept of Meta-Learning based on initialization together with FSL, this concept was related to the FSL technique and both contributed to improving the statistical accuracy metric average in image classification. The concept based on Meta-Learning metrics was also explored through the Siamese network and triple Siamese network algorithms. With the results obtained, the increase in accuracy was notable, especially when the models were trained using the Reptile algorithm together with the FSL technique on a set of similar images. The best results were obtained at an image resolution of 26 cm/px. This resolution, when used in the DenseNet121 model optimized with FSL and Reptile, achieved an accuracy of 81.3%, showing effectiveness when compared with the same architecture of the standard DenseNet121 without the use of such techniques. The Siamese network and triple Siamese network models also performed well, highlighting the importance of metric learning in the field of Meta-Learning. The results indicate a promising path for these methods, contributing to the construction of future artificial intelligence instruments that can help improve agricultural practices related to measuring productivity through images.