Biomass estimation and classification of secondary succession using radar and optical remote sensing data based on textural and spectral analysis in Amazonia.

摘要

Deforestation that happened during past decades in the Amazon Basin has been potentially affecting the global climate and carbon-cycle by contributing a vast amount of carbon into the atmosphere. Above-ground biomass (AGB) estimation of secondary successional and mature forests using remote sensing technology has been attracting scientists' attention. Optical sensor data are sensitive to forest canopies only, and therefore have limitations in AGB estimation, especially for tropical forests with complex structures and abundant species. Radar signals have the ability to penetrate canopies to detect the sub-layers of forests, and therefore potentially have a better performance in AGB estimation.; Based on the analysis of radar systems' operation mode, this study hypothesizes that radar backscatter variations in the range are more correlated to forest AGB than the variations in other directions. Developing a method to extract oriented high-frequency variations from radar images is the main purpose of this study. Non-scaling Wavelet Transformation (NSWT) was conceptualized to analyze such textual information in radar images.; Experiments were conducted in two study areas Rondonia and Bragantina using both atmosphere-calibrated Landsat TM data and JERS-1 L band radar data. The results were analyzed and compared between each other. The final results reveal that TM band 5 and 4 were the best representatives for estimating forest AGB. However, changes in structures and species combinations can disturb the negative relationship found between AGB and TM spectral values because forest spectral reflectance was significantly impacted by shadows in canopies. Dry season radar data performed well in AGB estimation. Its wavelet coefficients (W1ds) that represent high-frequency variations in the range direction were obviously more correlated to AGB than the textual information oriented in other directions. This result validates the hypothesis that radar backscatter variations in the range direction are more related to forest structures than the variations in other directions.; This study has also demonstrated that NSWT is helpful for improving land use/land cover classification accuracy in the Amazon Basin. NSWT can successfully remove high-frequency spectral heterogeneities within classes, e.g., shadows in forests, while keeping middle-frequency textual information that has proved useful in image classification.