Redes neurais artificiais aplicadas na identificação automática de áreas cafeeiras em imagens de satélite

Abstract

Coffee production is of fundamental importance in the southern region of Minas Gerais state and techniques for estimating the area planted are being intensely investigated in order to produce reliable yield estimates. Coffee trees have a similar spectral pattern to forest, making it difficult to automatically classify these land use types. The application of Artificial Neural Networks (ANN) to the classification of remotely sensed data is a promising approach in discriminating more complex classes of land use/cover. This study presents an application of Artificial Neural Networks for the automatic classification of remote sensing data which was adapted for the identification of land use with emphasis in the coffee areas in the municipalities of Guaxupé, Machado and Três Pontas, South region of Minas Gerais state in Brazil. The software used for preprocessing the data and validate the ratings was SPRING. To process and classify the images was used the IDRISI software. For the three regions the methodology for the application of ANN was divided in three main stages: in the first, the pre-processing of the data was accomplished. Masks were created in each one of the three bands of the images; in the second stage, the ANN was trained with representative samples of the classes of interest (coffee, forest, water, other uses and urban area) and applied on the image in order to verify its potential for discriminating established land use classes; the third stage consisted of the analysis and validation of the results, using as reference the map classified visually. The Kappa index was used to access the accuracy of the thematic maps obtained with the ANN. For the Guaxupé region, the Kappa index (k) was 71.85%, which is considered good. For the Machado region, the result of the classification by ANN was 55.84% for the more undulated part of the experimental area and 60.29% for the smoother landscapes. Although the kappa for Machado were lower than Guaxupé, the results obtained were better than what is found in the literature. For the region of Três Pontas the index Kappa obtained was 67,61%, that is also considered a good index. The ANN methodology presented good results. However, as the spectral bands only were not sufficient to carry out an optimized classification, it is recommended to use other environmental attributes as input data to improve the performance of the ANN.Keywords : Artificial Neural Networks, automatic classification, satellite image, coffee crop.