A multi‑decadal analysis of Pyroclastic Density Currents and lahar facies dynamics at Mount Sinabung, Indonesia (1995 to 2024) using multi‑sensor remote sensing and machine learning

Fahmi Arif Kurnianto, Indarto, Maulana Garaudy Purnomo

Mount Sinabung’s 2010 reactivation after a long repose period initiated a severe landscape-scale transformation driven by primary eruptive forces and secondary eruptive material dynamics. To address atmospheric data gaps caused by persistent cloud cover that uniquely constrains tropical volcanic monitoring compared to cloud-free non-tropical environments, this study implements a cloud-computing architecture within Google Earth Engine (GEE) using a 30-year multi-sensor dataset (1995–2024). Active microwave sensing via Sentinel-1 SAR is leveraged as the primary tool to map syn-eruptive Pyroclastic Density Currents (PDC) and post-eruptive lahar facies (debris flows and hyperconcentrated flows) in near real-time, while multi-decadal optical archives from landsat and Sentinel-2 are strategically utilized to monitor broader ecological and structural impacts due to PDC. Multi-temporal index analyses using delta Normalized Difference Vegetation Index (ΔNDVI) and delta Normalized Burn Ratio (ΔNBR) show long-term canopy degradation and the spatial trajectory of syn-eruptive PDCs, which systematically expanded towards the eastern and southeastern flanks along topographically steered valleys. For secondary material mapping, a Smile Random Forest classifier calibrated via a stratified 5-fold cross-validation framework achieved high classification performance with an Overall Accuracy (OA) between 99.71% and 99.82%, respectively. SAR backscatter analysis reveals a significant attenuation domain between −2 and −8 dB in ΔVV and ΔVH, representing the physical conversion from volume-scattering forest canopies into surface-scattering volcanic deposit tracks. Crucially, the spatial modelling indicates a seasonal hydroclimatic threshold during the December wet season peak under the west monsoon, which drives post-eruptive eruptive material remobilization. This study demonstrates that monitoring tropical volcanoes necessitates a dedicated multi-sensor framework where cloud-penetrating radar tracks immediate eruptive products and optical systems quantify multi-decadal environmental degradation, establishing a reproducible template for cascading hazard evaluation in dynamic volcanic arc settings.