Machine learning approach for peatland delineation using multi-sensor remote sensing data in Ogan Komering Ilir Regency

Abstract

Peatland is the region that is composed of accumulated layers of decayed vegetations and organic matters. It has an important role in regulating the hydrology cycle and maintaining global climate stability. Therefore protecting peatland from human intervention is imperative and the availability of accurate delineation map of peatland and non peatland is substantial and indispensable. Landsat 8 OLI and MODIS are proven to be beneficial for terrestrial analysis including peatland. For detecting peatland, the use of spectral bands to generate indexes that sensitive to the detection of Net Ecosystem Exchange (NEE) that related to peat water table, biomass vegetation and surface energy budget is the key. Therefore, several spectral-derived Landsat 8 OLI and MODIS products alongside with and DEM data were utilized. For classification, Machine Learning (ML) and Deep Learning (DL) methods namely Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGB) and Cat Boosting (CatB) were ML method that used alongside with Artificial Neural Network (ANN) of DL to delineate peatland distribution. The best result of the delineation was achieved by CatB algorithm with 75.52 % accuracy based on testing data and 82.61% based on validation data of field survey that held in Ogan Komering Ilir (OKI), South Sumatera Province.