Land Use Land Cover Classification using Deep Learning Techniques: A Comparative Study

Land Use and Land Cover (LULC) maps serve as an important component for natural resource monitoring. They convey insights into the utilization patterns of the Earth's land surface and the various forms of coverage present therein. In this research, we evaluated the performance of different deep learning architectures in the segmentation of satellite images into six distinct land cover classes such as trees, cropland, water bodies, built-up areas, roads and barren land. A unique dataset was curated utilizing Microsoft Bing Satellite imagery, encompassing the distinct land use classes within a region situated in the Mandya district, Karnataka. The digitization process was conducted using QGIS, facilitating the preparation of training and evaluation data for the development of trained Deep Learning (DL) models using popular DL architectures namely, FCN, U-Net, PSP, and DeeplabV3+. The performance of these models was compared with respect to their Fl score and mean Intersection of Union(mloU) metrics, on test datasets. From the comparison, it was observed that U-Net with ResNet-34 as the backbone demonstrated the best performance with a Mean IoU score of 0.803 and an overall Fl score of 0.89 on the test set. The generalization capability of the models was also evaluated by applying them to larger geographical areas, demonstrating their potential for Land Use Land Cover mapping for large areas.