Integration of Sentinel-1 and Sentinel-2 Data with the G-SMOTE Technique for Boosting Land Cover Classification Accuracy

The importance of Land Cover (LC) classification is recognized by an increasing number scholars who employ LC information in various applications (i.e., address global climate change and achieve sustainable development). However, studying the roles balancing data, image integration, performance different machine learning algorithms landscapes has not received as much attention from scientists. Therefore, present study investigates three frequently used Machine Learning (ML) algorithms, including Extreme Machines (ELM), Support Vector (SVM), Random Forest (RF) mapping at six landscapes. Moreover, Geometric Synthetic Minority Over-sampling Technique (G-SMOTE) was adopted to deal with class imbalance problem. In this work, time-series Sentinel-1 Sentinel-2 data were integrated improve accuracy, taking advantage both data. Machine-Recursive Feature Elimination (SVM-RFE) implemented distinguish most informative features. Based on results, RF G-SMOTE showed best result for four (coastal, cropland, desert, semi-arid). SVM had highest accuracy remaining two (plain mountain). Applied ML good performances landscapes, ranging Overall Accuracy (OA) 85% 93% RF, 83% 94% SVM, 84% 92% ELM. outcomes exhibit that although applying may slightly decrease OA values, it generally boosts results accuracies particularly minority classes.