Eruption Forecasting Using Random Forest on Single-Component Seismic Data: Insights from Three Indonesian Volcanoes (Semeru, Lewotobi Laki-laki, and Ruang)

This research employs a Random Forest machine learning method to forecast eruption probability for three Indonesian volcanoes: Semeru, Lewotobi Laki-laki, and Ruang. The primary objectives are to (1) evaluate model performance under varying data quality conditions and (2) test the transferability of forecasting models between different volcanic systems. We compare three scenarios: Semeru's major eruptions (December 2020 and 2021) with significant data gaps, Lewotobi Laki-laki's seven major eruptions (March - August 2025) with high data completeness, and Ruang's eruption (April 2024).

Seismic data from the vertical component (Z) were processed using Real-time Seismic Amplitude Measurement (RSAM), Displacement Seismic Amplitude Ratio (DSAR), and MSNoise to monitor seismic amplitude variations and relative velocity changes. Statistical methods extracted an initial set of 768 features from these processed signals. After removing highly correlated features, the top 20 most relevant features were selected for model training.

For Semeru, a model trained on the 2020 eruption successfully forecasted the 2021 eruption, with forecast probability exceeding the 0.7 threshold 12 hours prior to the eruption. For Lewotobi Laki-laki, models trained on earlier eruptions (March-April 2025) successfully forecasted subsequent event in May 2025, achieving lead times ranging from 6 hours to 1 day. Cross-volcano testing revealed that the Semeru-trained model failed to forecast the Ruang eruption, likely due to data incompleteness. In contrast, the Lewotobi Laki-laki model successfully forecasted the Ruang eruption 6 hours in advance, demonstrating successful model transferability. These results highlight the critical importance of data completeness for developing robust, transferable eruption forecasting systems.