A Downscaling Algorithm for Snow Cover Extent Over the Tibetan Plateau Based on a Similar Conditional Probability and Otsu’s Method

Abstract

Constrained by the limitations of remote sensing data in terms of temporal resolution, spatial resolution, and time series availability, there is currently a lack of effective long-term, high-spatiotemporal-resolution snow cover extent (SCE) products for studying snow cover changes. In this article, an SCE downscaling algorithm named SCPOT, which integrates a similar conditional probability (SCP) and Otsu’s method, is proposed. The algorithm is tested based on Moderate Resolution Imaging Spectroradiometer (MODIS) SCE and Advanced Very High Resolution Radiometer (AVHRR) SCE over the Tibetan Plateau. The SCP is defined as the probability that two pixels at corresponding positions across different scales correspond to the same snow conditions, and Otsu’s method is used to classify images by finding the optimal threshold via maximizing the interclass variance. During SCPOT testing, the SCP between 500-m MODIS and 5-km AVHRR SCEs from 2017 to 2018 was calculated, and the optimal segmentation thresholds for the SCP were determined via Otsu’s method. Then, based on the SCP and Otsu’s thresholds, the AVHRR SCE from 2015 to 2016 was downscaled to obtain 500-m resolution SCE, and the missing pixels were filled with space-time cubes and multivariate data. Evaluated with contemporaneous MODIS SCE and Landsat-8 SCE as reference data, the proposed downscaling algorithm has higher accuracy than nearest-neighbor resampling does, demonstrating feasibility in producing long-term SCE products with high spatiotemporal resolution via the algorithm.