Authors: Aaron Meneghini*, Clark University, Florencia Sangermano, Clark University
Topics: Remote Sensing
Keywords: Sentinel 1, Sentinel 2, Random Forest Classification, Remote Sensing, SAR, Tropical Remote Sensing
Session Type: Guided Poster
Start / End Time: 3:05 PM / 4:45 PM
Room: Roosevelt 3.5, Marriott, Exhibition Level
Presentation File: No File Uploaded
This study evaluates Sentinel-1 and Sentinel-2 remotely sensed images for tropical land cover classification. The two 10 meter backscatter bands of Sentinel-1 and four 10 meter optical bands of Sentinel-2 were used to create six land cover classifications across two study areas along the border of the Bolivian Pando Department and the Brazilian state of Rondônia. Data products were preprocessed through the open source software, Sentinel Application Platform (SNAP), and classified through a python based random forest classifier. Results indicate that Sentinel-2 optical bands possess a higher overall performance in delineating land cover types than the Sentinel-1 backscatter bands. The Sentinel-1 backscatter bands demonstrated the capability of delineating land cover types based on their textural properties but did not accurately separate similarly textured classes. The combination of Sentinel 1 and 2 results in slightly higher overall accuracy for delineating land cover types while increasing class homogeneity. Overall, the individual performance of Sentinel-1 and the increase seen in the land cover classifications derived from the data fused products is contrasted against the rigorous preprocessing requirements of radar backscatter data. While Sentinel-2 is demonstrated to consistently capture land cover in cloud free scenes, there is potential for Sentinel-1 backscatter bands to act as ancillary information in Sentinel-2 scenes affected by clouds. This use of Sentinel-1, would be particularly effective for studies focused on detecting classes with highly contrasting textural properties.