Use of artificial neural networks in the classification of degradation levels of pastures/Utilizacao de redes neurais artificiais na classificacao de niveis de degradacao em pastagens

The aim of this work is to evaluate the artificial neural networks and the maximum likelihood classification performances to classify land uses at Vicosa, Minas Gerais State, using ASTER images in order to verify degradation levels of pastures. In this study, three different levels of pasture degradation have been identified (moderate, strong and very strong) and an image composition of 3 bands was tested (covering the visible and the near infra-red) with 15 m of spatial resolution. The neural networks simulator used was the "Neural Java Network Simulator", with a feed forward model and the learning algorithm of back propagation. The obtained results show that the classification using neural networks, while presenting a slightly superior result, had a statistically similar performance compared to the maximum likelihood, getting a Kappa index of 0.80, against 0.79, respectively. In relation to individual performances, the class that presented the greatest error of classification was pasture in the level of very strong degradation, while the largest accuracy in the classification was obtained for coffee, for both classifiers, with 100 and 96% (respectively, Maxver and neural networks). Key words: aster, remote sensing, supervised classification Este trabalho teve por objetivo avaliar a eficiencia dos classificadores redes neurais artificiais (RNA) e o de maxima verossimilhanca (Maxver) na classificacao do uso da terra no municipio de Vicosa, MG, a partir de imagens do sensor ASTER, com enfase nos niveis de degradacao das pastagens. Neste estudo, foram identificados tres niveis de degradacao das pastagens (moderado, forte e muito forte) e avaliada uma composicao da imagem do sensor ASTER contendo as 3 bandas do visivel e infravermelho proximo, com resolucao espacial de 15 m. O simulador de redes neurais empregado foi o "Java Neural Network Simulator" e o algoritmo de aprendizado, o backpropagation. Os resultados mostram que a classificacao por redes neurais, embora apresente resultado ligeiramente superior, teve desempenho estatisticamente semelhante ao obtido pela classificacao pelo Maxver, obtendo um indice Kappa de 0,80, contra 0,79, respectivamente. Nas classificacoes realizadas a classe que apresentou maior erro de classificacao foi a pastagem no nivel de degradacao forte, enquanto a maior exatidao na classificacao foi obtida pelo cafe, para ambos os classificadores, com 100 e 96%, respectivamente. Palavras-chave: aster, sensoriamento remoto, classificac