A knowledge-augmented deep fusion method for estimating near-surface air temperature

Abstract

Near-surface air temperature (Ta) is a critical meteorological variable, and obtaining its precise spatiotemporal distribution is essential for numerous scientific domains beyond meteorology and hydrology. Despite the promising advancements in Ta mapping using machine learning, these models often suffer from inadequate generalization capabilities due to their heavy reliance on data. A critical limitation is that their “free” learning style fails to deeply uncover the intricate spatiotemporal patterns of Ta. Addressing this problem, we propose a novel knowledge-augmented deep fusion method (KADF), designed to enhance the accuracy of Ta mapping through integrating prior knowledge. KADF integrates three categories of prior knowledge concerning Ta: spatial autocorrelation, temporal autocorrelation, and temporal heterogeneity in the relationship between Ta and predictive variables. This tailored strategy enables the model to more efficiently explore the intricate spatiotemporal relationships between grounded Ta observations and satellite-derived auxiliary variables, culminating in accurate Ta estimates through the deep fusion of these datasets. The efficacy of KADF was thoroughly evaluated over Chinese mainland. The validation results show that KADF accurately mapped the spatiotemporal distribution of daily Ta, with root mean square error (RMSE) values of 1.0 °C for mean Ta (T_mean), 1.22 °C for maximum Ta (T_max), and 1.33 °C for minimum Ta (T_min). Moreover, the integration of prior knowledge regarding Ta significantly enhanced the generalizability of the data-driven mapping model. Compared to the state-of-the-art machine learning-based estimation method, KADF reduced the mean absolute error (MAE) values by 23–31 % and RMSEs by 24–29 %. Furthermore, this method considerably improved the ability to capture spatial and temporal variations in Ta across various environmental conditions. Finally, a 1 km daily Ta dataset for the time frame spanning from 2010 to 2018 was produced. Overall, KADF holds great promise for accurately estimating Ta and can be easily adapted to other regions. The source code of KADF has been made publicly available at https://github.com/Henu-frch/KADF.