Characterizing Forest Degradation using Multiple SAR Approaches

Editor's Notes

• #2: Description: Picture: West Kalimantan tropical peatlands view from air borne
• #3: Description: This slide is to highlight CIFOR research on tropical wetlands carbon stocks and GHG emission measurement (SWAMP activities)
• #4: Description: This study focuses on peatland forest ecosystem at Kampar Peninsular, Sumatera, Indonesia, which is geographically located between 101°50’ E and 104°07’ E and between 0°10’ dan 1°14’ N (Figure 1) The Kampar Peninsula is a wetland forest consists of great numbers of natural resources. The area has a size of approximately 700,000 ha, and peatlands in this region are mostly covered with 2-layer forest canopies. Peatlands in this region have depth from 50 cm (very shallow) to more than 20 meter (very deep). The peatland forests in Kampar Peninsula cover 17% of total peatlands in Riau Province (4.044 million ha), and for the entire country this province covers 56% of total 20.5 million ha of peatlands in Indonesia [12].
• #5: Description: From sampling plots data, we estimate 114 Mg/ha of carbon above ground biomass in the peatland forests, while the first pioneer species dominated with Macaranga sp. has very low carbon stocks of 3.7 Mg/ha. Mangrove forests and rubber plantation have relatively similar carbon stocks of ~50 Mg/ha. Carbon AGB increases following the increase of stand volume, but the LAI does not correspond directly with the AGB (Table 1), as pioneer species in general has higher LAI than other sampled peat forest environments.
• #6: Description: One scene of quad-polarimetric ALOS PALSAR data acquired during the end of rain season on May 2010 and Landsat 5 Thematic Mapper (TM) from February 2010 are used to characterize forest degradation event. Calibration of radar backscatter and corrections to other topographic distortions were conducted. Noise filtering to reduce speckle effects is applied using Enhanced-Lee filter. Polarimetric features, i.e. entropy, anisotropy and alpha angle, were generated from the coherency matrices of polarimetric SAR data and plotted against training observation data over degraded peatlands, pristine peatland forest and other land uses to characterize the profiles of radar backscatter in discriminating different peat forests conditions. Polarimetric features: alpha angle (a), entropy (b) and anisotropy (c). Two additional polarimetric features were also calculated, PolSAR random volume over ground volume ratio (RVOG_mv) based on polarimetric data inversion and accumulation of polarimetric backscatter (span in decibel / span_db).
• #7: Description: We used α/H (alpha angle - entropy) plane to plot ground truth data identified from the SAR data and the figure shows that except for water body and cleared land which are dominated by flat surface scattering (single bounce scattering), the vegetated land cover classes mainly show volume scattering due to random scattering from tree canopy cover. This may partly explain that low carbon biomass represented by low/sparse vegetation has more surface scattering, whereas moderate/higher carbon biomass may be identified from their volume scattering behavior. However, from Figure 2 we could not assess if the application of both polarimetric features could better discriminate different vegetation classes as many of the training data points are falling in the region 5, which characterize the domination of volume scattering mechanism. It has been already well-accepted that vegetation cover represents mainly volume scattering behavior, therefore to better classify different vegetation cover conditions, the use of radar backscatter (sigma nought) and other ancillary data are highly recommended.
• #8: Description: This study found that MLP-NN outperformed MLC technique to classify different forest classes over tropical peatland ecosystem. We also find that for MLP-NN to achieve the highest classification accuracy, addition of redundant input on the classification model will not give contribute to improved accuracy. Whereas, the MLC technique has a trend of improved classification accuracy following the number of input data used in the classification. The formal agreement is that statistical classifier, such as MLC method may suffer from non-normal data distribution, however this study does not seem to show such cases. For MLP-NN classification, feature selection of input data that are useful for the classification would benefit for increasing the classification accuracy. These facts are unfortunately beyond the scope of the present study, and are interesting to further explore.
• #9: Description: This study shows an attempt to test multi SAR approaches for characterizing forest degradation event in tropical peat swamp forests. We found L-Band of quad-polarimetric SAR data may provide useful information to characterize such event. Additionally, the accuracy of SAR classification depends mainly from variations of input data used for the classification, as well as the classification techniques being used. Applying MLC method, the SAR classification yielded 77.8% of accuracy combining radar backscatter, polarimetric features, RVOG_mv and span_db. While, MLP-NN classification is more sensitive to redundant input data, therefore the highest accuracy (79.9%) was achieved applying radar backscatter and multi-spectral Landsat bands in the classification.
• #10: Description:
• #11: Description: