Extraction of oil palm plantations on the undulating terrains in the Borneo using PALSAR Global Mosaic

Conversions of forests and peat swamps into oil palm plantations might cause decrease of net ecosystem production, change of water stream and loss of biodiversity. Most of the oil plantations in the South East Asia have been expanded in the past decade. For monitoring the distribution and condition of these plantations, the PALSAR Global Mosaic data set (PGM) have been utilized. This PGM is one of the L-Band Synthetic Aparture Radar (SAR) data sets orthorectified and mosaicked (unified). Unlike optical satellite imagery, the L-band SAR is useful especially for cloudy tropic regions. In addition, PGM have high resolution (about 10 m) and contains cross polarization (HV) SAR data which is useful for observation of forest because cross polarization SAR data contain data of the volume scattering which reflect the volume of plant bodies. On the other hand, topographic effect in PGM is not reduced sufficiently because of low resolution of DEM utilized to make PGM. As a result, pixel value of PGM is affected by the highlight effect especially in undulating terrains. These undulating terrains consist of rises which have about 10m height, exist in about 100m horizontal interval and cause striped patterns on SAR images. These patterns result in difficulty in extracting oil palm plantation using SAR imagery in the undulating terrains. However, many papers extracting oil palm plantations didn't show the accuracy of distributions extracted as oil palm plantation in the undulating terrains and thus it isn't clear how well oil palm plantations on the undulating terrains can be extracted. In this study, we carried out a supervised classification and extracted oil palm plantations in the north-west of Borneo Island. The Island is a part of the South East Asia and contains undulating terrains. In this extraction, we used the PGM data, a learning model and the training data made from PGM data, aerial photograph, high resolution optical satellite data and field survey data. After extraction, we assessed accuracy of the area extracted as oil palm plantations on undulating terrains using land cover maps made from high resolution aerial photo graph, optical satellite imagery and field survey data. In addition, we discussed the distribution of each land covers in a feature space (scatter diagram for many types of values such as pixel value and filtered pixel value which belong to each land covers) and discussed the structure of the learning model for extraction. As a result, we confirmed that the oil palm plantations on the undulating terrains can be extracted with high accuracy when PGM is used.