Assessing Performance of Convolutional Features for Terrain Classification Using Remote Sensing Data

Abstract

Advancements in satellite imagery and image processing techniques has enabled computerized solutions for various geographical and ecological monitoring and analysis problems. This has assisted geographical experts in topographic mapping, disaster monitoring and analysis of urban sprawl. Automated terrain segmentation and classification is a significant preliminary step in any Geographic Information System (GIS) based application. Nevertheless it is the most challenging task as well. In this paper, we present a complete schematic for assessing the effectiveness of various features and classifiers that are popularly employed in the literature for landform classification. A subset of images from DeepSat SAT-6 dataset is reconstructed using the RGB spectral band information and evaluations are performed using all six terrain classes. A number of color and texture features are then extracted from samples of each class and fed to a number of classifiers. In addition to conventional features, the effectiveness of convolutional features for terrain classification is also assessed in this study. A light convolutional neural network is proposed and trained on the employed dataset. The study also highlights the effect of color, texture and convolutional features on classification of each type of terrain under consideration. Experimental results show that convolutional features outperform both texture and color features by achieveing an overall accuracy of 93%.