Mintegration of C-band SAR and Multi-Spectral Data for Mapping Tropical Forest Areas in Southeast Asia
Waske, Björn1; Pflugmacher, Dirk2; Widayati, Atiek3; Budiman, Arif4; Hostert, Patrick2
1University Bonn, GERMANY; 2Humboldt University Berlin, GERMANY; 3World Agroforestry Centre, INDONESIA; 4WWF Indonesia, INDONESIA
Monitoring of tropical deforestation and forest degradation is required for ongoing climate change mitigation efforts such as REDD+. However, mapping tropical forest cover and associated land-uses with multi-spectral data alone is challenging as persistent cloud and aerosol contamination limits the availability of high-quality optical satellite imagery. SAR sensors have the ability to penetrate clouds and the backscatter is related to forest structural characteristics. SAR backscatter is also influenced by variable environmental conditions throughout the year and consequently the of time series is useful. Besides the use of multitemporal data, more robust classifications are usually achieved via data fusion multispectral data. To accurately map tropical areas, where forests are fast-growing and form complex mosaics with other land cover, will require frequent, high spatial resolution observations. In insular Southeast Asia, canopy openings can close within 1-2 years after disturbance. Secondary forest regrowth is often mixed with smallholder rubber, fruit gardens and fallow fields from shifting cultivation. A clear separation of these land-use types is difficult, due to spectral ambiguities, particularly when using monotemporal data from a single (optical) EO systems. One potential means to improve classification and change detection is by fusion of multiple SAR acquisitions and multi-spectral data. The integration of SAR with multi-spectral data has been useful in several land cover classification studies. It remains to be tested if such an approach would also help with mapping tropical deforestation and forest degradation. The specific objective of this study was the enhanced mapping of a humid tropical site in Borneo, Indonesia, by fusion of C-band SAR time series and Landsat data. We used ENVISAT ASAR and Landsat data from 2007 and 2008. However, conventional statistical classifier methods are often limited in context of multitemporal and multisensor data sets, because often multisource imagery cannot be modeled by a convenient multivariate statistical model. Moreover, the high-dimensional feature space of multi-source data may decrease classification accuracy with conventional classifiers when the number of training samples is small. Thus more advanced classification methods, taken from the field of machine learning are more adequate. In this study we used Support Vector Machines and Random Forests for data fusion and classification. These methods have been proven to be adequate for classifying multisensor and multitemporal data sets, even with limited number of training samples. Besides the use of multisensor and multitemporal data sets the integration of spatial information in the classification process seems interesting. Therefore, image segmentation was performed, using an in-house MATLAB implementation of the so-called Superpixel Contour (SPc) segmentation algorithm. Because of the long acquisition record of Landsat and C-band radar (ERS and ENVISAT ASAR), results of this study will also be useful for historic analyses of land-use changes. In light of the Landsat LDCM and Sentinel-1/2 missions, the successful integration of C-band SAR and Landsat data has promising implications for continuing operational tropical forest monitoring.