A Computer Vision-based Wetland Classification and Runoff Delineation Approach for Discontinuous Permafrost Landscapes of Northern Canada

Permafrost thaw as a product of climate warming is transforming permafrost plateaus to wetlands in the Canadian north. This effects the hydrological function of wetland-dominated watersheds. To be able to assess the impact of climate change in discontinuous permafrost areas, we require classified imagery which discriminates between forested areas, isolated wetlands, and connected wetlands. The classified imagery can be used for long-term change detection and as an input into hydrological models which require information about wetland connectivity and contributing areas.

A combination of high resolution RGB imagery (as a product of remote sensing) and deep learning methods is here used for land cover classification. We incorporated advance computer vision techniques to devise an automated and fast solution for:1: classifying dominant land covers in discontinuous permafrost zones. 2: discriminating between isolated and connected wetlands. 3: delineating the contributing area of both isolated and connected wetlands.

First, we trained a semantic segmentation model (with ResNet convolutional neural network architecture) using only the RGB bands of a WorldView-2 Satellite imagery of wetland-dominated portions of the southern Taiga plains in the Northwest Territories (NWT) to classify RGB imagery into wetlands, lakes, and forested area. A second processing phase categorizes the classified wetlands into isolated or connected wetlands based on their geometric connectivity, as determined by a multi-layer perceptron (MLP) algorithm. This avoids the common classification pitfall of trying to distinguish between (visually similar) isolated bogs and fens using a pixel-by-pixel classification scheme. Finally, we used the watershed algorithms to delineate estimates of the primary and secondary contributing areas of the categorized wetlands generated in phase two; this supports further analysis on hydrological effects of land cover transition over time. We applied the models to 10 study sites in the NWT and compared to manual classification. The semantic segmentation method reached an accuracy of 97 percent in classification of water, forested area, and wetland; the accuracy of MLP model for detecting isolated wetlands from other types of wetlands was 98 percent.