Wetland Classification Using Simulated NISAR Data: a case study in Louisiana

Identifying wetland’s spatial distribution is essential for their restoration and management due their significant role in the global water and carbon cycles. This study aims to assess the ability of upcoming NISAR data for delineating similar wetland classes using machine learning techniques. In particular, we investigated the synergistic use of several polarimetric features for efficient classification of wetland types. To this end, 84 polarimetric features from 11 polarimetric decompositions were extracted from full-polarimetry simulated NISAR data. The mean-shift algorithm was employed to segment the imagery for importing to the object-based machine learning classifiers. Post-classification feature importance analysis using the Gini index suggests that H/A/ALPHA, Freeman-Durden, and Aghababaee decomposition parameters have the highest contribution to the overall accuracy. Further, overall accuracies of 74.33% and 81.93% obtained by SVM and RF, respectively, demonstrated a great capability of NISAR data for wetland mapping and monitoring using limited available training data. Overall, the proposed approach for manipulating the upcoming NISAR data will provide some insight on the efficiency of an upcoming trend in using multi-frequency and full-polarimetry NISAR data with Sweep-SAR architect for producing high-resolution land cover maps on a global scale.