Rock and snow differentiation from colour (RGB) images

Abstract. We present a new method for differentiating snow and rock in colour imagery for application (including by remote sensing non-specialists) to multidisciplinary geospatial analyses in the Polar Regions (e.g. glaciology, geology, and biology). Existing methods for differentiating rock from snow and ice for land cover analysis in the Polar Regions rely on infrared or near-infrared imagery (e.g. the Normalised Difference Snow Index, NDSI). However, colour images are more abundant and higher resolution. To enable application of this resource, we present and review supervised and unsupervised methods for differentiating rock and snow from colour images. Whilst the unsupervised methods (fuzzy membership and a normalised difference index) are unable to accurately differentiate snow and rock from colour images, supervised classification (Maximum Likelihood Classification (MLC) and a new approach, Polynomial Thresholding (PT)) do achieve high classification accuracies (95 ± 2 % for PT and 94 ± 3 % for MLC, compared with manual delineation). The greater user control of PT achieves better accuracies than MLC in shaded areas (a challenge in high latitudes) and less extensive outcrops. We present the workflow for the new PT method, and provide a calibration tool for its implementation. This approach improves the possible resolution of Polar land cover analysis, and the increases the volume of data that can be utilised.