A Tile-Based Approach for the LULC Classification of Sentinel Image Using Deep Learning Techniques

Abstract

In this paper, we present a tile-based approach for the classification of sentinel images. Google Earth Engine provides open and free access to the sentinel level-2 and other satellite images. The study area of this research includes Bangalore (BBMP limits) of Karnataka state, India. We have created novel dataset by tiling sentinel image and obtained at least 1000 training samples for each of the five classes namely Forest, Open land, Water, Urban and Vegetation. Deep learning models such as VGG16, DenseNet and ResNet50 are used. Out of these three Resnet50 outperformed with classification accuracy of 98.47 on test data. All the image patches used here are of a spatial resolution of 8m. They are geo-referenced and manually labeled. This aids for exploring different applications of spatial data analytics.