Evaluating Wetland Mapping Techniques for New Brunswick Using Landsat-5 TM, ALOS-Palsar and Radarsat-2 Dual-Polarized Images
Friday, 19 December 2014
This study evaluates the use of dual-polarized (HH, HV) RADARSAT-2 C-band and ALOS-PALSAR L-band SAR images with LANDSAT-5 TM and a digital elevation model (DEM) for mapping wetland areas in New Brunswick. The resulting maps were compared to GPS field data that were collected in 2012 as well as to two wetland maps currently in use by the Province of New Brunswick, namely the Department of Natural Resources (DNR) wetland map and the forested wetland classes of the DNR forest map (called hereafter “DNR forested wetland map”). Overall the Random Forests classifier gave better classification accuracies than the maximum likelihood classifier. The comparison with the 146 wetland ground truth sites shows that 73.3% are correctly identified using the LANDSAT-5 TM classified image. For the SAR-based classified images, the number of correctly identified wetland ground truth sites is higher when the image acquired during the flooding is considered, the difference being higher with the ALOS-PALSAR images than with the RADARSAT-2 images. The number of correctly identified ground truth wetland sites is the highest when both the ALOS-PALSAR images and RADARSAT-2 images are used (98.6%). This percentage is well above the one obtained with the DNR wetland and forested wetland maps (44.5%). For both SAR-based classifications, the majority of the misidentifications are due to wetlands not being classified in the right wetland class and very few are wetland sites being classified as a non-wetland class. For the DNR maps, about half of the misclassifications are field-validated wetlands that are not mapped as wetland on the DNR maps, the remaining half are wetland sites classified in the wrong wetland class.