Aerosol Index and Machine Learning for Fire Smoke Mapping using The Second-generation Global Imager (SGLI) Data over Tropical Peatland Environments

Abstract

In this study, we explored the second-generation global imager (SGLI) of Global Change Observation Mission-Climate (GCOM-C) data to map biomass fires over tropical peatlands in Indonesia. The Absorbing Aerosol Index (AAI) derived from the near-Ultraviolet spectrum of SGLI at 250 m spatial resolution was examined statistically to perform smoke and other aerosol sources mapping. The mean values of AAI were statistically different among smoke, cloud, and other aerosols; however, the histogram distribution of AAI over those objects suggested a mixture of AAI regions between smoke and cloud. Machine learning algorithms overcame this limitation. Random Forest (RF) algorithm performs better than the Support Vector Machine (SVM) in mapping smoke from the cloud and other aerosol sources using all bands of SGLI, including the nonpolarization bands, polarization bands, and AAI image. RF performs 87% overall accuracy in classifying four objects, i.e., smoke, cloud, other aerosols, and free-aerosol background objects. The RF accuracy increased to 97% in mapping two classes, i.e., smoke and non-smoke, with the error of commission and omission at 4% and 3%, respectively. This finding provides a high potential for using SGLI data by RF algorithm for smoke detection over tropical peatland regions. More training samples of smoke in various conditions can enrich the artificial intelligence smoke database, which can be adapted as the input for developing the RF modeling using other hyperspectral sensors.