Development of a cloud-free MODIS NDSI dataset (2001–2020) over Northeast China

In the MODIS snow product collection 6.1 (C6.1), some snowy pixels are misclassified as cloud pixels due to an excessive cloud mask algorithm. This increases cloud contamination in normalized difference snow index (NDSI) and affects snow mapping accuracy. Northeast China is used as an example to develop excessive cloud mask removal and gap-filling algorithms, and generate a daily cloud-free NDSI dataset for each snow season from 2001 to 2020. To identify pseudo-cloud pixels, the green band and NDSI were used to distinguish forested snow and clouds, thin snow in plains and clouds, respectively. This decreased cloud from nearly 60% to below 30%. We developed a spatiotemporal cube cloud removal algorithm based on NDSI similarity (STNSI) for the remaining clouds in which the central cloudy pixel was filled with neighborhood pixels in the spatiotemporal cube. Meanwhile, bias correction was performed, and the average cloud decreased to below 1% after the first iteration. The produced STNSI NDSI has high precision under different underlying surfaces, with OA and FS values greater than 0.9. Compared with the existing STAR NDSI, STNSI NDSI snow mapping accuracy improved, especially in forested areas where OA improved by 37%. Moreover, removing many pseudo-cloud pixels greatly enhances gap-filling efficiency.