Temporal Variability of SAR Backscatter over Tropical Regions in Tanzania,The Democratic Republic of Congo and Colombia
Kohling, Miguel; Haarpaintner, Joerg

This paper presents an initial study for the development of an automatized Synthetic Aperture Radar (SAR) remote sensing based monitoring system to support the Monitoring, Reporting and Verification (MRV) of REDD (Reducing Emissions from Deforestation and forest Degradation) activities. Within this work the temporal and spatial variability of SAR backscatter over tropical areas is analyzed. The seasonal signature trends of L-band and C-band SAR data from ALOS PALSAR, Envisat ASAR, RADARSAT-2 and ERS-2 SAR are studied for different and locally specific forest and other land cover types in Tanzania, the Democratic Republic of Congo (DRC) and Colombia. The localization of the study sites within important regions of the worldwide tropical rain forest allows for global comparison of our results and a presumption of their global transferability. The study comprises three steps: First, with focus on seasonal variability, the inner-class backscatter distribution functions are statistically analyzed over the available time series. The training polygons for feature extraction were delineated by local experts or, if not available, by interpretation of optical VHR (Very High Resolution) data. In case of a high inner-seasonal variability, e.g. due to strong precipitation events, temporal filtering of the backscattering data is required before further processing. In contrast, low variability allows the usage of spatially filtered single acquisitions. This question is of relevance, when considering the data usability in a real-time monitoring system. Second, mono-seasonal and bi-seasonal data sets are created by temporal filtering and are each classified by unsupervised classification. A comparison of those provides a general impression about class-separability and information gain with a multi-temporal approach. Finally, a supervised classification is conducted, based on the aforementioned training polygons, again on mono- and bi-seasonal data. The results of the unsupervised and supervised classifications are then compared in terms of class separability. This procedure can reveal important, previously neglected sub-classes, which would improve any further supervised classification process. Our results confirm that the analysis of SAR backscatter time-series is of tremendous importance for reliable classification of different land cover types over tropical areas. In addition, the class-specific backscatter signatures are summarized to training data sets for future supervised or model-based classification activities.