Snow Coverage Prediction using Machine Learning Techniques

Abstract

Snow coverage is often predicted through analysis of satellite images. Two of the most common satellites used for predictions are MODIS and Landsat. Unfortunately, snow coverage predictions are limited either by MODIS images sets' low resolution quality or Landsat dataset's low temporal frequency. In this study, we employed a set of various machine learning techniques, including multilayer perceptrons (MLP), random forest regressor (RF), and convolutional neural networks (CNN) to model the relationship between high temporal frequency of MODIS data and high spatial resolution of Landsat data. Through various experiments, we propose an improved Fractional Snow Coverage (FSC) based on relationship between RGB, lower frequency infrared channels and regional locality.