Development of a new forest snow mapping algorithm using MODIS data, machine learning and time-lapse photography

Remotely sensed MODIS (Moderate Resolution Imaging Spectroradiometer) data and the NDSI (Normalized Difference Snow Index) based approach have been applied globally for snow cover mapping. However, this method displays severe omission errors in forested areas, due to the forest canopy shading of snow. In this study, we developed a new forest snow mapping algorithm based on MODIS reflectance data, time-lapse observations of forest snow, and a random forest model. We built a time-lapse camera network in the eastern Qilian Mountains in northwestern China to monitor the forest snow processes and obtain the ground truth data. The random forest (RF) model seems to be powerful in capturing the relationships between the MODIS surface reflectance bands and the forest snow presence. The presented approach significantly improved the accuracy of binary snow cover (BSC) mapping in forests. We evaluated the performance of the proposed algorithm with the traditional NDSI-based method. The results show that the new algorithm has a superior performance in forest BSC mapping, compared to the NDSI-based BSC. The proposed RF-BSC can retrieve ~70% of all real forest snow pixels, while the NDSI-BSC can only detect 8-14%. We further investigated the geographical influence (e.g. topography, forest coverage, and solar illumination) of the algorithm performance. This study suggests that the fusion of optical remote sensing data and ground-based observations using machine learning techniques has a great potential in improving the accuracy of land cover mapping.