Snow detection in alpine regions with Convolutional Neural Networks: discriminating snow from cold clouds and water body. Issue 1 (31st December 2022)

• Record Type: Journal Article
• Authors: Lu, Yichen; James, Thomas; Schillaci, Calogero; Lipani, Aldo
• Is Part Of: GIScience & remote sensing. Volume 59:Issue 1(2022)
• Page Start: 1321
• Page End: 1343
• Publication Date: 2022-12-31
• Subjects: Snow -- Alps -- Convolutional Neural Networks -- mapping -- DeepLabV3+; Geodesy -- Periodicals; Cartography -- Periodicals; Aerial photogrammetry -- Periodicals; Remote sensing -- Periodicals; 526.05
• DOI: 10.1080/15481603.2022.2112391
• Languages: English
• ISSNs: 1548-1603

Abstract

ABSTRACT: Accurately monitoring the variation of snow cover from remote sensing is vital since it assists in various fields including prediction of floods, control of runoff values, and the ice regime of rivers. Spectral indices methods are traditional ways to realize snow segmentation, including the most common one – the Normalized Difference Snow Index (NDSI), which utilizes the combination of green and short-wave infrared (SWIR) bands. In addition, spectral indices methods heavily depend on the optimal threshold to determine the accuracy, making it time-consuming to find optimal values for different places. Convolutional neural networks ensemble model with DeepLabV3+ was employed as sub-models for snow segmentation using (Sentinel-2), which aims to distinguish clouds and water body from snow. The imagery dataset generated in this article contains sites in global alpine regions such as Tibetan Plateau in China, the Alps in Switzerland, Alaska in the United States, Southern Patagonian Icefield in Chile, Tsylos Provincial Park, Tatsamenie Peak, and Dalton Peak in Canada. To overcome the limitation of DeepLabV3+, which only accepts three channels as input features, and the need to use six features: green, red, blue, near-infraRed, SWIR, and NDSI, 20 three-channel DeepLabV3+ sub-models, were constructed with different combinations of three features and then ensembled together. The proposed ensemble model showed superior performance than benchmark spectral indices method, with mIoU values ranging from 0.8075 to 0.9538 in different test sites. The results of this project contribute to the development of automated snow segmentation tools to assist earth observation applications.