Page 109 - 《高原气象》2026年第3期
P. 109
3 期 王彩玲等:基于地形引导注意力的降水降尺度模型研究 717
降尺度预测模型[J]. 高原气象, 42(3): 646-656. DOI: 10. of summer precipitation in Northwest China[J]. Plateau Meteorol‐
7522/j. issn. 1000-0534. 2022. 00102. Zhu J T, Yang Q Y, Li X, ogy, 42(3): 646-656. DOI: 10. 7522/j. issn. 1000-0534. 2022.
et al, 2023. Characteristics and east downscaling forecast model 00102.
A Deep Learning-Based Precipitation Downscaling
Models with Terrain-Guided Attention
WANG Cailing , FAN Lei , XIE Xiaoning 2
1
1, 2
(1. School of Computer Science, Xi’an Shiyou University, Xi'an 710065, Shanxi, China;
2. State Key Laboratory of Loess Science, Institute of Earth Environment,
Chinese Academy of Sciences, Xi'an 710061, Shanxi, China)
Abstract: As a powerful deep-learning downscaling technique, convolutional neural networks (CNN) are wide‐
ly utilized to generate high-resolution precipitation data from low-resolution global climate models through data-
driven approaches, playing a critical role in assessing the impacts of climate change at both regional and local
scales. This study proposes a CNN-based precipitation downscaling model, the Topography-Guided Attention
Network (TGAN), which downscales coarse-resolution (2°) atmospheric variables to produce high-resolution
precipitation fields at 0. 1°. The model adopts a Laplacian pyramid as a progressive, multi-level downscaling
framework, in which the spatial resolution of precipitation fields is incrementally enhanced and precipitation
structures are reconstructed through successive stages. In addition, a topography-guided attention module is in‐
corporated, which leverages an attention mechanism to integrate atmospheric variables with elevation data at cor‐
responding spatial scales. By combining these multi-scale features, the module strengthens the network’s capaci‐
ty for feature representation and learning, thereby improving the accuracy and reliability of simulated precipita‐
tion. Focusing on the middle reaches of the Yellow River, TGAN is trained on daily ERA5 atmospheric vari‐
ables, GPM IMERG daily precipitation data from 2001 to 2010, together with static elevation data, and validat‐
ed using data from 2011 to 2020. The results indicate that, compared with a conventional CNN model, TGAN
achieves higher accuracy in spatiotemporal precipitation simulations at daily, monthly, and annual scales. At the
daily scale, TGAN achieves a lower average root mean square error (5. 10 mm·d ) and a higher average correla‐
-1
tion coefficient (0. 42) compared with the conventional CNN model. Additionally, TGAN more accurately cap‐
tures extreme precipitation events (95th and 99th percentiles) and better aligns with GPM IMERG observations
in probability density distribution, particularly for heavy precipitation range. This study further investigates the
impact of different loss functions on the downscaling performance of TGAN. Using the RMSE loss function im‐
proves overall predictive accuracy but leads to underestimation of extreme precipitation events, whereas the Ber‐
noulli-Gamma loss function, although slightly less accurate overall, more faithfully reproduces extreme precipi‐
tation events. Its probability density distributions are highly consistent with both GPM IMERG data and station
observations, indicating that the model has an enhanced capability to capture extreme precipitation events, there‐
by better reproducing the distribution characteristics of precipitation frequency. Overall, by combining the topog‐
raphy-guided attention mechanism with the Bernoulli-Gamma loss function, TGAN demonstrates clear advantag‐
es in downscaling precipitation over the middle reaches of the Yellow River, not only improving overall simula‐
tion accuracy but also better representing extreme precipitation events, providing a robust and reliable tool for
high-resolution precipitation modeling in complex terrain regions.
Key words: precipitation downscaling; convolutional neural network; terrain-guided attention; Laplacian Pyra‐
mid; Middle Reaches of the Yellow River

