Modeling Aboveground Biomass in Tropical Forests Using Multi-Frequency SAR Data-A Comparison of Methods

In the context of climate change mitigation mechanisms for avoiding deforestation, i.e., reducing emissions from deforestation and forest degradation (REDD+), comprehensive forest monitoring, especially in tropical regions, is of high relevance. A precise determination of forest carbon stocks or above-ground biomass (AGB) for large areas is of special importance. This study analyzes and compares three different methods for retrieving AGB in Indonesia's peat swamp forests from multi-frequency SAR backscatter data. Field inventory AGB data were related to LiDAR measurements allowing plentiful accurate AGB estimations. These estimatedAGB data provided a powerful basis for SAR based AGB model calibration and validation. Multivariate linear regression (MLR), artificial neural network (ANN) and support vector regression (SVR) were examined for their performance to retrieve AGB on the basis of multi-temporal TerraSAR-X and ALOS PALSAR imagery. The MLR model resulted in lower coefficients of determination and higher error measures than the other two approaches and showed significant overestimations in the high biomass range. The SVR modeled AGB was more accurate than ANN modeled AGB in terms of independent validation, but showed less variation in the spatial distribution of AGB and saturated at approximately 260 t/ha. The ANN model showed a superior performance for modeling AGB up to 650 t/ha without a saturation in the lower biomass ranges. For the needs of REDD+, it is very important to know the possibilities, constraints and uncertainties of AGB retrieval based on satellite imagery.