Automated monitoring of sub-decadal changes in glacier covered area using Google Earth Engine and a suite of ancillary datasets

Abstract:

As we continue to see rapid changes in glacier covered areas globally, there remains a need for automated approaches to monitor both the spatial and temporal extent of these changes. Understanding glacial change over broad spatial extents and at high temporal granularity are critical foundations for further research, monitoring and decision making.

Here we test the efficacy of a temporal segmentation and classification workflow to resolve the glacier covered area of Alaska at 2-4 year time steps and 30m spatial resolution. To do this we modified the LandTrendr algorithm, which is widely used to monitor land cover/land use change and disturbance, and applied the algorithm to all available minimum snow extent (late summer) optical Landsat images (Landsat 4 TM, Landsat 5 TM, Landsat 7 ETM+ and Landsat 8 OLI) over Alaska from 1984-present. LandTrendr leverages the petabyte scale data archive and compute resources of the Google Earth Engine platform to describe landscape change in a consistent manner over time. Subsequent classification is then carried out using the spatiotemporal exploratory model (STEM) framework, which allows for spatially-varying model relationships across broad geographic domains. Our methods also utilize ancillary datasets, including the Alaska 5m IFSAR elevation data and MODIS snow cover data in classification and post processing steps to mitigate challenging edge cases, such as late/early season snow and debris covered glaciers, which frequently undermine fully automated workflows. Our approach also creates multi-annual composites to maximize image availability, amends the native CFmask to address locations where snow/ice/cloud pixels are frequently conflated, and uses a snow season metrics dataset derived from MODIS to tune the temporal window of image collection. Based on a limited spatial and temporal extent reference dataset from the Randolph Glacier Inventory, we report promising (85-92%) initial accuracies in the Gulf of Alaska area based on our fully automated approach. Although currently limited to the regional scale, the approach holds promise for wider application in global glacier covered areas and may provide guidance for additional inventory and monitoring and change studies in the future.