A trained Mask R-CNN model over PlanetScope imagery for very-high resolution surface water mapping in boreal forest-tundra

• Published in: Remote Sensing of Environment
• Main Authors: Freitas, Pedro, Vieira, Gonçalo, Canário, João, Vincent, Warwick F., Pina, Pedro, Mora, Carla
• Format: Article in Journal/Newspaper
• Language: English
• Published: 2024
• Subjects: Mask R-CNN; Deep learning; Planet; Scope; Arctic and subarctic; Water mapping; Small water bodies; Arctic; Hudson Bay; Nunavik; Canada; Hudson; permafrost; Subarctic; Tundra
• Online Access: http://hdl.handle.net/10451/62838; https://doi.org/10.1016/j.rse.2024.114047

Small water bodies (< 0.01 km2) showing diverse limnological properties occur in great abundance across the boreal forest and tundra landscapes of the Arctic and Subarctic. However, their classification, geographical distribution and collective importance for water, heat, nutrient, contaminant and carbon cycles are still poorly constrained. One important step for better understanding the role and evolution of small water bodies in the fast-changing northern landscapes is to develop image analysis protocols that allow their automatic remote sensing detection, delineation and inventory. In this study, we set an image analysis protocol (High Latitude Water – HLWATER V1.0) based on a trained supervised Mask R-CNN deep learning model over PlanetScope imagery for the automatic detection and delineation of small lakes and ponds that were absent in existing datasets. Most of our training dataset comprised water bodies smaller than 0.01 km2 (97%) and spanned a wide range of environmental and hydrological settings, from the sporadic to the continuous permafrost zones of Canada. The model was tested as a fully autonomous approach for eastern Hudson Bay, Nunavik (Subarctic Canada), a region that poses challenges for water remote sensing given the abundance and variety of small water bodies. These are mainly permafrost thaw and glacial basin ponds in the boreal forest-tundra in challenging optical settings influenced by vegetation or topography shadowing, or revealing peat water logging, fen and bog pond conditions. A multi-scale validation approach was developed using water body delineations from PlanetScope imagery and ultra-high resolution orthomosaics from Unoccupied Aerial Systems. This procedure allowed a sub-pixel assessment and identified the limitations and strengths of the trained model for detecting small and large water bodies. The results varied according to different landscape units, with mean Intersection over Union (IoU) 0.5 F1 Scores of 0.53 to 0.71 and mean F1 Scores of 0.62 to 0.95. Considering 166 m2 ...