Land surface change detection after major volcanic eruptions in Indonesia using Machine learning and Spatial-Temporal transferability

• Historical eruption data enhances model accuracy in data-scarce environments; • Integration of SAR and optical data enhances volcanic deposit mapping; • Random Forest and XGBoost consistently outperform SVM for volcanic impact mapping. Mitigating volcanic eruption disasters is a key priority in Indonesia’s national development agenda. Due to the high density of active volcanoes near populated areas, even eruptions with a relatively low Volcanic Explosivity Index (VEI) can result in significant fatalities and losses. Therefore, rapid mapping of eruption impacts is essential both during and following an eruption. Machine learning with Earth Observation (EO) data enhances the speed and accuracy of volcanic impact mapping, critical for timely disaster response. In this study, we evaluate the performance of three machine learning classifiers − Random Forest, Support Vector Machine (SVM), XGBoost – for mapping land surface changes resulting from volcanic deposits (e.g. lahars, pyroclastic density currents and lava) across three different data availability scenarios using multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical data. Scenario I use a traditional approach, where training and validation data are sourced from the same eruption event. In Scenario II , the spatial transfer scenario, we simulate the absence of local training data by using data from different volcanoes. In Scenario III , the spatial–temporal transfer scenario, we include training data from past eruptions of the target volcano in addition to other events. Our results demonstrate that Scenario I achieve the highest classification accuracy, with Random Forest and XGBoost consistently outperforming SVM, achieving f1-scores exceeding 0.90. Scenario III improves average model performance compared to Scenario II (Δf1-score – 0.12), highlighting the value of historical eruption data for enhancing classification accuracy in data-scarce environments. These findings underscore the critical role of integrating EO data and machine learning in advancing volcanic impact mapping, offering a scalable and efficient approach to support disaster management efforts globally.