To better understand and mitigate threats to the long-term health and functioning of wetlands, there is need to establish comprehensive inventorying and monitoring programs. Here, remote sensing data and machine learning techniques that could support or substitute traditional field-based data collection are evaluated. For the Bay of Quinte on Lake Ontario, Canada, different combinations of multi-angle/temporal quad pol RADARSAT-2, simulated compact pol RADARSAT Constellation Mission (RCM), and high and low spatial resolution Digital Elevation and Surface Models (DEM and DSM, respectively) were used to classify six land cover classes with Random Forests: shallow water, marsh, swamp, water, forest, and agriculture/non-forested. Results demonstrate that high accuracies can be achieved with multi-temporal SAR data alone (e.g., user’s and producer’s accuracies ≥90% for a model based on a spring image and a summer image), or via fusion of SAR and DEM and DSM data for single dates/incidence angles (e.g., user’s and producer’s accuracies ≥90% for a model based on a spring image, DEM, and DSM data). For all models based on single SAR images, simulated compact pol data generally achieved lower accuracies than quad pol RADARSAT-2 data. However, it was possible to compensate for observed differences through either multi-temporal/angle data fusion or the inclusion of DEM and DSM data (i.e., as a result, there was not a statistically significant difference between multiple models). With a higher repeat-pass cycle than RADARSAT-2, RCM is expected to be a reliable source of C-band SAR data that will contribute positively to ongoing efforts to inventory wetlands and monitor change in areas containing the same land cover classes evaluated here.
Wetland Classification with Multi-Angle/Temporal SAR Using Random Forests