Estimating AVHRR snow cover fraction by coupling physical constraints into a deep learning framework

• Published in: ISPRS journal of photogrammetry and remote sensing Vol. 218; pp. 120 - 135
• Main Authors: Zhao, Qin, Hao, Xiaohua, Che, Tao, Shao, Donghang, Ji, Wenzheng, Luo, Siqiong, Huang, Guanghui, Feng, Tianwen, Dong, Leilei, Sun, Xingliang, Li, Hongyi, Wang, Jian
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2024
• Subjects: AVHRR; Deep learning; Northern hemisphere; Physical constraints; Snow cover fraction; Deep learning; Snow cover fraction; Physical constraints; Northern hemisphere; AVHRR
• ISSN: 0924-2716
• DOI: 10.1016/j.isprsjprs.2024.08.015

Accurate snow cover information is crucial for studying global climate and hydrology. Although deep learning has innovated snow cover fraction (SCF) retrieval, its effectiveness in practical application remains limited. This limitation stems from its reliance on appropriate training data and the necessity for more advanced interpretability. To overcome these challenges, a novel deep learning framework model by coupling the asymptotic radiative transfer (ART) model was developed to retrieve the Northern Hemisphere SCF based on advanced very high-resolution radiometer (AVHRR) surface reflectance data, named the ART-DL SCF model. Using Landsat 5 snow cover images as the reference SCF, the new model incorporates snow surface albedo retrieval from the ART model as a physical constraint into relevant snow identification parameters. Comprehensive validation results with Landsat reference SCF show an RMSE of 0.2228, an NMAD of 0.1227, and a bias of −0.0013. Moreover, the binary validation reveals an overall accuracy of 90.20%, with omission and commission errors both below 10%. Significantly, introducing physical constraints both improves the accuracy and stability of the model and mitigates underestimation issues. Compared to the model without physical constraints, the ART-DL SCF model shows a marked reduction of 4.79 percentage points in the RMSE and 5.35 percentage points in MAE. These accuracies were significantly higher than the currently available SnowCCI AVHRR products from the European Space Agency (ESA). Additionally, the model exhibits strong temporal and spatial generalizability and performs well in forest areas. This study presents a physical model coupled with deep learning for SCF retrieval that can better serve global climatic, hydrological, and other related studies.