Peatland Delineation Using Remote Sensing Data in Sumatera Island

Abstract

Peatland is important for global climate as it stores enormous amount of carbon and if degraded could yield devastating effect to atmosphere as it worsen the Earth’s green house level. Given that fact, it is inevitable to start conserving peatland area. The conservation plan requires reliable and clear delineation map to differ between peatland and non peatland. Remote sensing technology is effective tool to solve this task. Its recent products such as Landsat, MODIS and ASTER GDEM are potentially capable of identifying and characterizing peatland. Employing spectral analysis make it possible to identify peatland unique features and discriminate between peat area and non – peat area. Machine Learning (ML) method was used to produce peatland map as it was able to identify class signature data with high dimensionality feature. From early assessment, ML was able to perform classification with accuracy more than 80% using solely testing and training dataset from South Sumatera province. By only using the knowledge from training data in South Sumatera, ML classified Riau, Jambi and South Sumatera itself. The result was quite promising as accuracy attained by Random Forest and Gradient Boosting were 79.95% and 78.60% for 500 meter spacing grid training data, and 73.95% and 78.10% for 750 meter spacing grid training data. The use of machine learning in remote sensing for classification despite not providing perfect result can still be a useful tool to give an insight to solve highly complex classification task.