Comparative Analysis of two Ga- based Classification Algorithms over heterogeneous areas.

The identification on land cover is complex because of fast changes in heterogeneous areas. To cover large regions classification algorithms have been developed based on the exploitation of satellite images. Most of the existing have used optical images. More studies recent works have also included PolSAR data.

However, these algorithms have been validated ever urban and/or agricultural areas with classes clearly identified, and only few studies have focused on heterogeneous regions.

Among different classification algorithms these based on GA-ANN and GA-SVM have shown better performance. In this study we compare the performance of these two approaches for urban areas, agricultural lands, and a heterogeneous tropical forest. The urban area is the San Francisco city, the agricultural land by a corn production area in Central Mexico, and the heterogeneous forest by a tropical forest in south area Mexico.

Particularly, this study aims at identifying the most sensitive parameters from optical images, C band polarimeters, SAR discriminators, and textural parameters.

The training kappa shown that the most sensitive parameters for the urban areas come from the coherence and covariance matrix, the parametric decomposition and the texture parameters of the hh, hv and vv polarization with/or accuracy of 97%. considering four classes.

For the agricultural area in the classification the most sensitive parameters are going by the information of the scattering, covariance and coherence matrices, the decomposition of Freeman-Durden and Cloude-Pottier, the parametric decomposition and texture parameters with/or accuracy of 96% considering five classes and, for the tropical forest, the most sensitive parameters are provided by the SAR discriminators, the information of the covariance and coherence matrices, the decomposition of Freeman-Durden and Cloude-Pottier and texture parameters present an essential role in the classification., with/or accuracy of 95%, considering four clases.

Finally, the performance of the algorithms over the complete images is currently implemented. Preliminary results indicate a reduction in the kappa value of about 20%.

Keywords: SAR Data, LandSat-7, LandSat-8, RadarSat-2, Land Cover, Genetic algorithm (GA), Artificial Neural Network (ANN), Support Vector Machine (SVM).