An Automatic Data-Driven Framework for AVHRR Long-Term Fractional Snow Cover Retrieval using Dynamic High-accuracy Training Samples from Landsat

Snow cover dynamics in the High Mountain Asia (HMA) region serve as key indicators of global climate change. Although Fractional Snow Cover (FSC) products derived from MODIS data have advanced significantly, their relatively short temporal span limits long-term snow trend analyses. In contrast, the Advanced Very High Resolution Radiometer (AVHRR) offers the longest continuous record of optical remote sensing data, presenting a unique opportunity to capture interannual variability in snow cover across decades. However, given the relatively limited observational capability of AVHRR, conventional physically based methods face considerable challenges in achieving accurate retrievals. To address this, we propose a two-stage, data-driven FSC retrieval approach for the HMA region, employing Random Forest (RF) and XGBoost models separately in the regression stage and a RF in the classification stage to correct the results. Based on this approach, two FSC products (AVHRR FSC L1) spanning 1984–2018 were developed, incorporating cloud indicators. FSC L1 datasets were validated with Landsat 8 FSC and in situ snow depth, and benchmarked against the SnowCCI AVHRR products from the European Space Agency (ESA FSC). The results demonstrate clear improvements: the FSC L1 datasets achieve RMSE values of 0.079 (XGBoost) and 0.082 (RF), with corresponding R² values of 0.955 and 0.953, respectively, outperforming the ESA FSC (RMSE 0.097, R² 0.934). These models demonstrate superior accuracy in forested and topographically complex areas, as well as improved consistency in spatial snow distribution. To further enhance data usability, we introduce a multi-step spatiotemporal gap-filling algorithm and produce a cloud-free FSC product (AVHRR FSC L2). Its performance shows RMSE 0.084 (high cloud cover) and 0.154 (medium–low cloud cover), with R² 0.862 and 0.919. A key innovation of this study lies in the automatic construction of an extensive, high-quality training dataset on the Google Earth Engine (GEE) cloud platform, enabling the first AVHRR-based, long-term, high-accuracy FSC dataset specifically tailored for the HMA region. The resulting FSC products are expected to support a broad range of applications in cryospheric research, climate change monitoring, hydrological modeling, and regional water resource management.