Lava flow mapping Karangetang Volcano during 2019 eruption using Sentinel-2 Images and Random Forest model

Abstract

Abstract Karangetang is one of Indonesia’s most active volcanoes, located in northern Sulawesi. At least 200 people were evacuated due to the threat of lava during the nearly year-long eruption of 2019. In this study, we aim to map the lava flow using Sentinel-2 images. We used a random forest algorithm to separate lava and non-lava classes. Approximately 402 training points were visually interpreted from a pair of pre- and post-event images in 2019 classified as cloud, cloud shadow, bare land, settlement, vegetation, recent lava, and historical lava. To begin, we employed all bands and spectral indices previously identified as useful for separating lava from other materials. A region growing algorithm was used along with additional input data (temperature anomalies product, Normalized Hotspot Indices, and digital terrain model) to differentiate recent and old lava. Only the significant variables from the first run were kept that had a cumulative contribution to the model of greater than 10%. The top 6 important features are Digital Terrain Model, bands 11 and 12 (SWIR) of post-image and pre-image, and band 8A (NIR) of post-image. In the cross-validation of the Random Forest classification, our preliminary results indicate an accuracy of 97% for the test data (N-tree=100). The region’s growing algorithm helped distinguish between recent and historical lava. In the future, we intend to validate the map using an independent dataset and test the model at various locations