A unique Ruleset for Saltmarsh and Mangrove monitoring using Sentinel-2

Abstract

Mangrove and saltmarsh vegetation are two important communities in the wetland ecosystem that require continuous monitoring considering their ecological threats and status. Remote Sensing observation is one of the tools to monitor this community. However algorithms are also important for the best sensor to map it accurately. Although there is some research about eCognition-based mapping, a unique rule set is important to monitor this community. Considering this research gap, a unique threshold-based ruleset has been generated in the Random Forest Model environment. A range of vegetation indices, individual bands of Sentinel 2 and Digital Elevation Model (DEM) data were tested in the feature selection method to find the best features for a Random Forest (RF) model. Top three variables were selected and those are (a) reNDVI_A, (b) VIRE_A, and (c) SWIR1 which gave 98.37% accuracy for the test data. A similar trend was found for the other three sites when they were compared with observed data. For site 1, Mangrove was 84.73%, Saltmarsh was 60.19%, and Mixed was 70.12% accurate. A similar trend was found for other three sites with overall accuracy 87.74 % for site-2, 66.23% for site 3, and 72.98% for site 4. A unique ruleset for wetland extent estimation will help to map a wide range of areas with a very minimum level of fieldwork. It will also reduce the cost of mapping done by visual observation, measurement and extensive fieldwork. The results of this work will provide the necessary insight and motivation for ecologists, and environmentalists.