Chapter 2 - Machine learning for snow cover mapping

Abstract

Optical remote sensing data has long been used to map snow cover from space. However, the available spatial and temporal resolution is often insufficient to accurately monitor changes in snow cover, particularly across complex terrain. More recently, Planet small satellites have provided new opportunities to map daily snow cover at approximately 3-m spatial resolution, with more than 200 in-orbit satellites. In this chapter, we present a workflow that uses a random forest (RF) model to map snow-covered areas (SCA) from Planet imagery. We demonstrate how to prepare the model inputs, optimize model parameters to obtain the best results, and evaluate model performance using a reference dataset derived from the Airborne Snow Observatory (ASO) snow depth dataset. The RF model achieves promising performance, with an overall accuracy of 89% and an F1 score of 0.87. The model performs better in open areas (90%) than in forested areas (85%). The RF model is lightweight and easy to set up, making it feasible for use in future applications. Given its good performance in SCA mapping, it can be successfully applied to address water resource management, snow modeling, and prediction problems. In conclusion, our work demonstrates the use of RF models for mapping snow cover, which provides a valuable contribution to the field and can serve as a useful guide for future applications.