Land use and Land Cover Classification for Dang District Nepal using Satellite Imagery and Machine Learning on Google Earth Engine

Abstract

Land use land cover classification (LULC) is a key tool for accessing, monitoring and management of natural resources. Advanced remote sensing technologies such as satellite imageries and machine learning algorithms have been widely used for LULC classification around the globe. This study was aimed to compare and analyze the performance of Random Forest (RF) and Classification and Regression Tree (CART) algorithms for LULC classification of Dang district using Landsat-9 and Sentinel-2 imageries of the year 2023 on Google Earth Engine (GEE) platform.  During the study, satellite images were accessed and filtered by predetermined region of interest, date, cloud percentage (<10%) and spatial resolution (30m) followed by cloud masking and median composite. Several satellite indices including normalized difference vegetation index (NDVI), normalized difference built up index (NDBI), modified normalized difference water index (MNDWI) and barren soil index (BSI) were computed and used to detect five different LULC classes i.e., crop lands, water bodies, forest and shrubs, settlements, and barren and sandy lands. The CART model classified the Landsat-9 and Sentinel-2 imageries more accurately with overall accuracies of 97.43% and 96.41% as compared to RF model i.e., 95.71% and 95.71 respectively. Similarly, the Kappa coefficient for CART was 0.97 for Landsat-9 and 0.95 for Sentinel-2 imageries while that of RF was 0.94 for both image sources. The results indicate that CART performed comparatively better than RF under same level of resolution. The current study suggests government stakeholders and policymakers to employ LULC as major key tool for sustainable land management, ecological conservation, and socio-economic development and addressing global concerns such as urbanization and climate change.