Classificação automática de áreas cafeeiras em imagens de satélite, utilizando Redes Neurais Artificiais

Abstract

Digital satellite imagery classification in recent decades has been increasingly recognized as a technique for breaching the shortages of thematic mapping. Artificial Neural Networks (ANN) are currently an emerging option for sorting data from different sources. Minas Gerais state stands out from other Brazilian states in terms of planted area and quantity of coffee produced. As environmental factors have a great influence on the quality of the drink, an understanding of the environment in which the coffee is inserted allows for better production planning, aiming at sustainable development, and the achievement of quality products. Thus, the objective of this work is create an algorithm using the software MATLAB toolbox and Neural Networks for classification of this coffee areas in satellite images, to identify and map the use and occupation of land, with an emphasis on identifying coffee areas Três Pontas of the region located in the south of Minas Gerais. To implement and utilize the MatlabNeural Networks Toolbox software we use convolutional ANN, that is, a multilayer perceptron specifically designed to recognize two-dimensional shapes. We also use the backpropagation algorithm for neural network training. The project will be carried out in various stages: problem definition, pre and post-processing of data, data collection training and testing, network structure design, training, testing and validation. In this study, we demonstrated the potential of neural networks in automatic classification of coffee areas in satellite images, which justifies further work in this area. The Kappa index of accuracy obtained in the coffee region of Três Pontas was 0.6711, with an overall rate of 79.6025%, areas of coffee had a hit of 76.25%. Future work aimed at using other types of images, resulting from other satellites and other regions, besides the addition of other features like shape, texture, vegetation index, etc.