Identificação de extensas áreas de culturas agrícolas empregando uma abordagem espectro-temporal utilizando imagens MODIS

• Autor(a) principal: Braga, Alessandra Lopes
• Data de Publicação: 2007
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: LOCUS Repositório Institucional da UFV
• Texto Completo: http://locus.ufv.br/handle/123456789/3846
• topic: Classificação multitemporal; MODIS; Redes neurais artificiais; Imagens multiespectrais; Remote sensing; MODIS; Artificial neural networks; Multispectral photography; CNPQ::CIENCIAS EXATAS E DA TERRA::GEOCIENCIAS::GEOFISICA::SENSORIAMENTO REMOTO

Remote sensing images have spatial, spectral, radiometric and temporal characteristics, which become an important tool for agricultural applications, in several aspects. This thesis describes a methodology for classification of extensive agricultural areas, in regional scale, using MODIS (Moderate Resolution Imaging Spectro-radiometer) images. It was used a spectral-temporal surface response, where each pixel of the image is represented in a three-dimensional space and the axes are: time, wavelength and reflectance, respectively. The methodology consists of an interpolation of analytical surfaces, passing through control points, using two types of interpolators (Collocation and Polynomial Trend Surfaces). Through these interpolation methods, the surface coefficients were generated, which describe the distribution of the surface in the three-dimensional space. These coefficients were then used into the classification process. Two classification algorithms were used, the maxima likelihood and artificial neural network classifiers. One of the drawbacks, in supervised classification process, is the acquisition of the reference image. For this work were used three distinct methodologies for its attainment: thematic image sampling from the spatial resolution of 30m to 250m; digitalization of homogeneous polygons on the screen; and neighborhood operation, which consists of the elimination of pixels with neighborhood with high variation on the digital level. Statistical analyses were used in order to validate the results. The results show that the classification using neural networks had the best resulted, even with small training sample size. The results also show the importance of high quality reference image generations.