Assessing the Influence of Temporal Compositing on Land Cover Classification Accuracy Using Multi-Source Sentinel Data in Google Earth Engine

Land cover mapping is essential for environmental monitoring and spatial planning, while recent advances in cloud-based platforms such as Google Earth Engine (GEE), as well as the increasing availability of free Earth Observation data, have enabled efficient and scalable classification workflows. Although numerous studies have focused on comparing machine learning algorithms, few have systematically evaluated the influence of temporal compositing strategies on land cover classification accuracy using a consistent methodological framework. To isolate the influence of temporal information, all experiments were performed using an identical classification framework. This study investigates the impact of temporal compositing strategies on land cover classification accuracy in the Suceava River Basin (Romania) using a processing workflow based on multi-source Sentinel-1 SAR and Sentinel-2 optical imagery together with percentile-based temporal descriptors (10th, 25th, 50th, 75th, and 95th percentiles) derived from spectral bands and spectral indices (NDVI, NDWI, and NDBI). By maintaining the same Random Forest classifier and processing workflow throughout the analysis, the study isolates the effect of temporal compositing on classification performance. Three temporal scenarios were evaluated for 2025: (i) annual composite, (ii) vegetation season composite (April–September), and (iii) summer composite (June–August). Land cover classification was performed using 1,524 reference samples, with 70% used for training and 30% for validation. High classification accuracies were obtained for all scenarios, with Overall Accuracy (OA) values of 96.52%, 96.73%, and 97.14% and corresponding Kappa coefficients of 0.958, 0.961, and 0.966 for the annual, vegetation season, and summer composites, respectively. Despite relying on substantially fewer input images, the summer composite achieved the highest classification accuracy, demonstrating that the selection of an appropriate temporal window can improve classification efficiency without compromising accuracy. The proposed workflow provides a transferable framework for operational land cover mapping and highlights the importance of temporal compositing strategies in multi-source remote sensing applications.