PEMODELAN SPASIAL TINGKAT KERENTANAN KEBAKARAN MULTI KEDALAMAN LAHAN GAMBUT TROPIS BERBASIS MACHINE LEARNING

• Main Author: Fadlansyah Ramadhan, Lukman
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69092
• Institution: Institut Teknologi Bandung

Tropical peatlands play a vital role as global carbon sinks, habitat for endangered flora & fauna, and provider of freshwater and source of livelihood for the people. Yet, their well-being and existence is currently threatened by increasingly severe and frequent wildfire linked to climate change. Therefore, quantifying peat vulnerability to fires and understanding the underlying driving mechanisms is crucial to developing mitigation and adaptation strategies. One such effort to accomplish that is by applying Machine Learning algorithm to spatially model multi-depth peatland fire vulnerability. Three algorithms used in this study are Random Forest, Classification and Regression Tree (CART), and Support Vector Machine (SVM). Data processing is done on Google Earth Engine. Fire data from MODIS satellite serves as the response variable, while the predictor variables consist of drought index, soil moisture, precipitation, and soil water & carbon content at six different depth layers. At model fitting stage, spatial block cross-validation is conducted iteratively five times yielding probability and binary map of peatland fire vulnerability. The study found that 1) Peatlands in Central Borneo, coastal South Sumatra, and parts of Riau and West Borneo are very vulnerable to fire and require utmost attention, 2) the most potent predictor variable in the models varies between soil carbon content and drought index/precipitation, and 3) output model of RF and CART fulfills AUC-ROC, Sensitivity-Specificity, and Overall Accuracy test, while SVM only fulfills the Overall Accuracy test.