Mapping snow cover in forests using optical remote sensing, machine learning and time-lapse photography

The accurate spatial information of snow cover is useful for understanding the impact of global warming, and it is of high significance for hydrological disaster prediction, water resources management, and climate change research. The Normalized Difference Snow Index (NDSI) based approach has been used extensively around the world for mapping snow, and they displayed high accuracy in open areas. However, capturing snow cover in forests remains problematic due to the obstruction effects of the forest canopy, which causes the snow cover area to be seriously underestimated. In this paper, we present a new algorithm based on machine learning (ML) technology to improve the accuracy of binary snow cover (BSC) mapping in forests, using the remotely sensed surface reflectance and ground truth data. A time-lapse photography network with a two-hour resolution was established in the eastern Qilian Mountains in northwestern China to obtain the ground truth data both in forests and open areas. We trained Random Forests (RF) with the 500-m Moderate Resolution Imaging Spectroradiometer (MODIS) reflectance data from bands 1–7 to generate BSC results (RF-BSC). Then we evaluated RF-BSC and the NDSI-derived BSC maps with three different NDSI thresholds (i.e., 0.10, 0.29, and 0.40) against ground-truth data. The results indicate that the proposed algorithm has a high performance in forest BSC mapping in this area, compared to the NDSI-threshold approach. The RF-BSC can retrieve 67% of all real forest snow pixels, while the NDSI-based BSC can only detect 8–14%. We also find that the performance of the algorithm seems to be sensitive to changes in solar illumination conditions and forest coverage. This study suggests that machine learning with the fusion of optical remote sensing and ground-based observations is an effective approach for improving the accuracy of forest snow cover mapping at regional scales.