Leveraging Machine Learning Algorithms and Sentinel-2 Satellite Imagery for Land Use Land Cover Classification of Bangladesh Using Google Earth Engine

Land Use Land Cover (LULC) classification for a developing country like Bangladesh faces considerable challenges like limited quality data and computational resources and the complexities of processing vast amounts of data. A suitable method to classify the large geographic area of Bangladesh remains understudied, implying a lack of information about the landscape of Bangladesh. To bridge this gap, we aim to identify the best classification technique specifically suited for the whole Bangladesh territory. We present a methodological approach that facilitates the efficient processing of Sentinel-2 imagery within Google Earth Engine (GEE) by incorporating a novel combination of spectral and topographic features of Bangladesh. We evaluated Gradient Tree Boost (GTB) along with the most commonly used Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART) algorithms using only a limited amount of training samples to classify our study region into waterbody, wetland, agriculture, forest, bareland, and built-up areas. To evaluate the classifiers, we used a set of assessment metrics including Overall Accuracy (OA), Kappa, Producer’s Accuracy (PA), and Consumer Accuracy (UA). Our approach obtained excellent accuracy for Random Forest with OA 93.8% and Kappa 0.92. The findings demonstrate that our method of leveraging machine learning techniques and GEE has the potential to overcome existing complexities in handling large-scale LULC classification for Bangladesh.