Comparative assessment of ensemble machine learning and Google Earth Engine for flood susceptibility mapping in the Dipota River Basin, Assam, India

Abstract Flooding poses a persistent threat to socio-economic stability and environmental sustainability in Northeast India, particularly within dynamic alluvial river systems. This study is an integrated flood susceptibility assessment framework combining ensemble machine learning (ML) models with a Google Earth Engine (GEE)-based screening approach for the data-scarce Dipota River Basin of Assam, where limited flood inventories and sparse hydrological observations limit conventional modelling. Multi-source topographic, hydro-climatic and land-use datasets were analysed using cloud-based geospatial processing and statistical learning techniques. Seven machine learning (ML) classifiers were trained using field-verified flood occurrence data and evaluated using threshold-dependent and threshold-independent metrics, followed by ensemble construction. Results indicate that Random Forest and Support Vector Machines achieved high predictive accuracy but showed tendencies toward overfitting, whereas CART, MaxEnt and MARS demonstrated more stable generalization under limited data conditions. Ensemble models improved spatial robustness, identifying approximately 59–79 km² classified as highly flood-susceptible. In contrast, the GEE-based framework delineated a markedly smaller high-risk extent (12.28 km²), reflecting differences between probabilistic learning and deterministic rule-based mapping. The study demonstrates that ensemble machine learning (ML) approaches are more effective for long-term flood susceptibility assessment in data-limited basins, while GEE offers rapid and operationally efficient hazard screening. The framework is transferable to flood-prone, data-scarce regions and supports risk-informed planning and disaster management.