Multi-temporal Sar and Optical Data Fusion with Texture Measures for Land Cover Classification Based on the Bayesian Theory

This paper addresses the land cover classification capabilities of multi-temporal synthetic aperture radar (SAR) data and optical data fusion based on Bayesian approach. Multi-temporal SAR data were used to extract average backscattering coefficient, backscatter temporal variability and long-term coherence while the reflectance values were calculated using the optical data. Grey Level Cooccurrence Matrix (GLCM) based texture measure including mean, standard deviation, correlation, contrast, homogeneity, dissimilarity, and entropy was used to parameterize texture in the image. These features are integrated in the Bayesian approach. Three processing steps for the classification were used in this study: 1) Information fission by feature extraction. 2) Supervised classification with greyscale value of information fission. Then, the maximum a posteriori (MAP) estimation can be used to label each class. 3) The combination of logical operators was applied to compute the final combined Bayesian membership value function. Finally, the classification results were generated taking Osaka city of Japan as the study area. In the experiment, fourteen ALOS/PALSAR level 1.1, single-polarization data, and ALOS/AVNIR-2 level 1B2G data were used. The major classes were selected to be built-up areas, fields, forests, and water bodies. It was found in SAR and optical images that mean images produced the best result among the texture measures because of the smoothing effect for image. Moreover, the correlation results of texture measurements with mean texture showed that using highly correlated textures can have lower accuracies. In addition to this, a slight increase in accuracy was found when using multi-temporal SAR and optical data fusion combining all texture images. The study shows that combination of various textures can improve over single-set texture, but the low correlation between textures should be considered. * Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author.