Enhanced multi-region satellite image classification accuracy in noisy environments using a hybrid statistical approach

Abstract

Classifying satellite images is complicated by the presence of Gaussian noise, which corrupts pixel relations and decreases the quality of the input data. In this research, an advanced hybrid statistical technique to classify satellite images into three classes: Urban, Vegetation, and Water is developed. The presented technique uses a two-step processing chain: the first step applies a median filter to reduce the effect of Gaussian noise; the next step combines the benefits of GLCM and LBP features for accurate texture classification. Testing of the developed algorithm was carried out using Sentinel-2 satellite images of Barcelona for different values of noise (from 0% to 20%). Experimental results show that the presented algorithm provides a classification accuracy of 98.55%. Unlike the existing GLCM and LBP approaches, which do not work correctly in the presence of noise, the presented one successfully eliminates bias in the area distribution caused by noise.