Journal Article

STARNet: A Deep‐Learning Algorithm for Surface Shortwave Radiation Retrieval From Fengyun‐4A

Shi, Hongrong, Liu, Bai, Ding, Zijing, Xia, Xiang'ao, Yang, Dazhi, Chen, Yun, Song, Mengmeng, Shen, Yanbo

2025

Overview

Satellite‐retrieved surface shortwave radiation is indispensable to solar energy meteorology applications. In stark contrast to conventional irradiance retrieval algorithms that are confined to individual pixel information, this work proposes the STARNet (Spatio‐Temporal Association‐based Retrieval Network), which is a deep‐learning algorithm that exploits the information embedded in the spatio‐temporal neighbors of a target pixel. The algorithm holds three technical innovations: (a) a data preprocessing method that highlights the correlation‐ and causality‐type climatology associations in the original reflectance and brightness temperature observations; (b) a graph network cascade that extracts topological spatio‐temporal features, and (c) a multi‐scale convolution network that extracts regular spatio‐temporal features. The empirical part of this work showcases irradiance retrieval from Fengyun‐4A over China. True out‐of‐sample verification demonstrates that STARNet can outperform physical and conventional data‐driven retrieval algorithms. Most importantly, STARNet is exceedingly general and thus applicable to many other retrieval tasks, such as those for aerosols or clouds. Plain Language Summary Gridded surface solar radiation cannot be directly observed but must be retrieved from top‐of‐the‐atmosphere reflectance images taken by satellites. Conventional pixel‐level retrieval methodologies are particularly vulnerable to atmospheric misrepresentation, as exemplified by the paradigmatic case where cloud presence in the sun‐to‐surface path fails to coincide with cloud presence in the surface‐to‐satellite path. To remedy these limitations, we advance a novel deep‐learning architecture that systematically incorporates spatio‐temporal correlations. The algorithm synergistically integrates several smaller neural networks (e.g., graph network and convolution network), each having a designed functionality that facilitates better feature extraction. Particular emphasis is placed on the optimal representation of long‐term climatological associations embedded within the multi‐dimensional data structure (encompassing spatial, temporal, and spectral dimensions). A case study with Fengyun‐4A data reveals that the proposed algorithm can outperform all competing methods considered. The ultimate deliverable constitutes a high‐accuracy surface radiation product with a 4‐km resolution over China. Key Points A 4‐km surface shortwave radiation product over China is developed based on Fengyun‐4A using deep learning Extracting topological and regular spatio‐temporal features using graph and convolution networks improves retrieval accuracy The proposed retrieval technique outperforms conventional machine‐ and deep‐learning models, as well as physical algorithms

• Publisher: John Wiley & Sons, Inc,Wiley
• Subject: Algorithms / Brightness temperature / climatological association / Climatology / Deep learning / Fengyun‐4A / Irradiance / Machine learning / Meteorology / multi‐scale regular spatiotemporal feature / Pixels / Radiation / Reflectance / Retrieval / satellite retrieval / Short wave radiation / Solar energy / Surface radiation temperature / surface shortwave radiation / Temporal variability / topological spatiotemporal feature