The potential for large changes in the amount and distribution of surface water in arctic landscapes is high given climate-induced changes in permafrost. Because changes in surface water have broadscale implications in the structure and function of ecosystems, understanding and tracking surface water change is a high priority for some Alaskan refuges. As such, the primary objective of this project is to build an end-to-end, scalable remote sensing infrastructure allowing for the efficient production of cloud-free, high-resolution composite imagery with associated surface water estimates for refuge lands across Alaska. To accomplish this goal, the Google Earth Engine platform is used, allowing for an automated process for querying, filtering, and extracting water pixels from dense stacks of multi-temporal and multispectral, 10-meter resolution imagery from the European Space Agency’s Sentinel-2 mission. The team employs a cloud detection algorithm to mask out cloudy pixels from all available images and achieve a cloud-free composite for a given timestep and over an entire region. Surface water features are then extracted using an automatic image thresholding approach based on the Normalized Difference Water Index (NDWI) calculated from each composite’s green and near-infrared (NIR) bands. To mitigate misclassification of persistent burned areas and topographic shading as water, the process automatically masked out areas in the resulting surface water extraction that are identified as either severe burns from a Normalized Burn Ratio (NBR) calculated from each composite’s NIR and short-wave infrared (SWIR) bands or areas representing steep topography as observed in an IfSAR-derived slope map. The algorithm returns a cloud-free color-infrared image composite and final surface water extraction for each timestep specified by the user.
To better understand the uncertainty of extracted surface water outputs, they are validated against ground truth water features digitized from high-resolution, co-occurring airborne imagery (downsampled to match target resolution) in a 4.7-square-kilometer region of Alaska’s Tetlin National Wildlife Refuge (NWR). This initial validation shows a precision of 97%, recall of 81%, intersection over union (IOU) of 79%, and overall pixel accuracy of 97%. The current methodology appears to be excluding waterbodies smaller than 2 acres. Future efforts may employ Sentinel-1 SAR imagery to increase the amount of usable imagery and higher resolution commercial datasets for observations of finer scale surface water dynamics. An immediate priority is to evaluate the spatiotemporal changes of surface water estimates observed within Tetlin NWR as an initial case study. This work may provide an operational workflow and product that provides surface water estimates at temporal and spatial resolutions useful for current and future management and modeling efforts.
Google Earth Engine outputs displaying: (Top left) 2016 Sentinel-2 color-infrared summer image composite in Tetlin NWR; (Top right) automatic surface water extraction for 2016 summer composite; (Bottom) number of summers with surface water occurrence between 2016 and2020.