Abstract
Snow cover is the most prevalent land cover type during northern hemisphere winter and thus a key component of the cryosphere, a major driver of Earth's climate system. Monitoring the variability of snow cover is vital to assess its impact on both global and regional climates and water resources. Since most snow-covered areas are in remote or inaccessible environments, where field measurements are sparse, spaceborne remote sensing, offering both frequent observations and global coverage, is widely used to monitor snow cover. In this study, we develop a deep convolutional encoder-decoder neural network to accurately distinguish snow from clouds and other land cover types over large geographical areas. The network is trained using two published collections of multispectral imagery from the Landsat-8 and Sentinel-2 satellites. When only using the visible blue, green, and red spectral bands, the network can classify snow cover with an F-score of 0.73 (0.7) for the Landsat-8 (Sentinel-2) dataset. Adding the near-infrared band increases the F-score to 0.79 (0.72) and adding the near-infrared and both shortwave-infrared spectral bands results in the best F-score of 0.82 (0.74).
The generalization ability of the network to unseen data, which is different but related to the training data, is assessed by training the network on the Sentinel-2 dataset and applying it to the Landsat-8 dataset, and vice versa. The network pretrained on the Sentinel-2 (Landsat-8) dataset shows an F-score of 0.74 (0.77) on the Landsat-8 (Sentinel-2) dataset, without applying any transfer learning or post-processing methods. When the network weights are additionally fine-tuned using supervised domain adaptation, the F-score is 0.8 (0.77). Overall, the F-scores indicate that the deep convolutional neural network can adapt to various image acquisition conditions over different geographical areas and can generalize to unseen data from a different sensor than the training data. The overall accuracy of the network is 92% (94%) on the Landsat-8 (Sentinel-2) dataset. In conclusion, the convolutional encoder-decoder network is a flexible end-to-end deep learning algorithm able to distinguish snow cover from other land cover types, which does not require complicated pre-processing or any auxiliary data sources.