Driving factors of formation of top-down gravelization encroachment in alpine hillslopes grasslands

Long-term Landsat series satellite images were used as data to monitor the dynamic changes of sloped grassland in all study sites. Landsat images with a 16-day revisit comprise seven multi-spectral bands with spatial resolution of 30 m (Senf et al., 2020). We hand-selected annual Landsat images closest to the phenological peak (defined as the mid-July to mid- August) and with little to no cloud cover in the study region (Piao et al., 2012). We screened United States Geological Survey (https://earthexplorer.usgs.gov/) archive to maximize time coverage, obtaining 9 to 20 images in four study areas from 1985 to 2020 (Table 2). The earliest image was 3rd of June and the latest image was 29th of September. The Digital Elevation Model (DEM) with spatial resolution of 12.5 m was downloaded from the United States Geological Survey, and the slope data derived from it.

Grassland cover fraction mapping and validation

The regression-based spectral unmixing approach has been demonstrated to be a powerful technique for extracting surface materials at a sub-pixel scale (Pacheco et al., 2010; Senf et al., 2020). To apply this technique, the images were used to create a spectral library with pure cover type spectra representative of the study area. Pure pixels were defined as unchanged pixel representing a type of cover (grassland, gravel, glacier or water) during all the years, and surrounded by pure pixels to avoid errors from spatial mis-registration. Overall, we collected 300 pure spectra as training data in each study area and surrounding area. The training data sets were used to train a fully constrained least squares (FCLS) unmixing model predicting the coverage from the mixed Landsat spectra. The trained model was finally applied to each Landsat image, continuously predicting grassland cover fractions throughout all 30 years.
High spatial resolution remote sensing images in Google Earth (GE) provides available way for validation of unmixing model. To acquire images from GE, we bought the copyright of Bigemap (http://www.bigemap.com/), which is a downloader of satellite image in GE. Eight remote sensing images (0.25-0.5 m) were downloaded for validation of unmixing results. Table 3 lists the locations, acquisition dates, and spatial resolutions of those GE images. We validated the model predictions by randomly sampling 50 reference Landsat pixels across years of GE images correspond in each study area, by comparing coverage in Landsat and GE images.