Exploiting Feature Extraction Techniques for Remote Sensing Image Classification

Multispectral image classification derived from satellite sensors is a topic of graet interest for the scientific community. The great interest is to automatically identify different areas including coffee production. The coffee stands out for being an important source of income and jobs, as well as being one of the most important products of the economy of Brazil. However, automatically map this culture has been a challenge so much for object-oriented analysis how much to methods based on "pixel to pixel" techniques. This work exploits different feature extraction techniques aiming at identifying the most discriminative features for remote image classification. The satellite image used in this study refers to the Tres Pontas region, Minas Gerais, Brazil, which has a great agricultural production, especially coffee. It has been used the seven spectral image bands of Landsat 8 OLI (Operational Land Imager). It was considered 5 land use classes: Coffee, Wood, Water, Urban Area, Other Uses (Grassland, Soil, Weathered, Other Cultures, Eucalyptus). Various spectral and textural characteristics were extracted as features and combined for the classification. Higher-order statistics-based features were also extracted and combined with those commonly used in the literature for remote sensing image classification. Two feature selection methods for dimention redution was used: the Fisher's Discriminant Ratio (FDR) and the linear correlation. As classifier, a multilayer perceptron has been used. The best Kappa indices obtained was 73.13% for the model that considered all extracted features (a total of 43) as input.