Sistema de classificação de imagens de sensores remotos

Abstract

The classification multispectral image derived from satellite sensors is an objective desired by the scientific community for identifying areas in multispectral images. The wide interest is in the identification of many areas, including coffee production. Coffee is highlighted as an important source of income and employment, as well as one of the most important products Brazilian economy . However, automatically mapping this culture has been a challenge both for object-oriented and for ”pixel to pixel” analyses. The objective of this work consists in the use of different parameters selecting methods, in order to identify the best parameters for the classification. The satellite image used in this study concerns the region of Três Pontas-Minas Gerais (MG), Brazil, which presents a strong agricultural production, especially of coffee. The chosen images were used in all seven spectral bands of the Landsat 8-OLI (Operational Land Imager). In the proposed methodology, we selected 5 classes of use: coffee, forest, water, urban area, other uses (pasture, exposed soil, other cultures, eucalypt). Many spectral traits and textures were combined for the classification. Trait performance based on higher order statistics (HOS) were also tested in combination with those commonly used in the literature for this end. Seven spectral bands of Landsat 8 satellite were explored. Two selection methods of dimension reduction were used: Fisher Discriminant Ratio (FDR) and linear correlation. The results showed that performances for both classifiers, ANN and SVM, were similar . The best kappa indexes obtained were 73.13% for ANN, fed by all traits extracted, and of 74.37% for SVM, by the pixels of the seven bands as input. The results were compared with the methods currently found in literature and with the commercial software.