Abstract
The Group on Earth Observations Global Network for Observations and Information in Mountain Environments (GEO-Mountains) is a GEO Work Programme Initiative that seeks to connect and facilitate access to diverse sources of mountain observation data and information at different scales. At a recent GEO-Mountains workshop, experts convened and identified ecosystem extent and fragmentation as key priority Essential Biodiversity Variables (EBVs) to monitor processes that account for and improve the understanding of ecological and social-ecological changes in mountain ecosystems. Monitoring these EBVs requires an accurate mapping of the distribution of different ecosystems and how these are affected by natural and human-caused disturbances. For this purpose, Sayre et al. (2020) developed a global map of terrestrial ecosystems based on land cover, terrain, and climate information at a spatial resolution of 250m. However, its utility for regional or local scale applications was not yet investigated due to (1) the lack of a thorough validation and (2) the relatively coarse spatial resolution, which may preclude its use in mountain environments. To address these two problems, we derive maps of mountain ecosystems at a spatial resolution of 30m for the Central European Alps and the Western Himalayas using artificial intelligence in combination with the latest optical remote sensing data. We follow and compare two different approaches to derive these maps for the years 2016-2020, thereby enabling a comprehensive assessment of ecosystem distribution and change in two diverse mountain environments:
* Directly downscaling the existing 250m terrestrial ecosystem map
* Downscaling the land cover information and following the approach of Sayre et al. (2020) Both downscaling approaches are based on machine learning algorithms using the Harmonized Landsat Sentinel-2 (HLS) surface reflectance dataset (Claverie et al. (2018)) in combination with digital elevation model and climate data. We focus on (1) the automatic generation of training data by filtering the existing coarse ecosystem information using the multispectral HLS remote sensing data and (2) comparing the advanced feature extraction capabilities of deep machine learning algorithms (e.g., convolutional neural networks) with traditional feature-based machine learning algorithms (e.g., random forests, support vector machines). The algorithms are implemented in the Microsoft Azure High Performance Computing (HPC) cloud platform, where the computationally expensive deep learning algorithms can make use of multiple Graphics Processing Units (GPUs) for large-scale image analysis.