当前位置：
X-MOL 学术
›
Atmos. Res.
›
论文详情
Our official English website, www.x-mol.net, welcomes your
feedback! (Note: you will need to create a separate account there.)
A deep learning framework for enhanced retrieval of atmospheric temperature and humidity profiles across China: Unifying inversion algorithms across multiple stations
Atmospheric Research ( IF 4.5 ) Pub Date : 2024-11-17 , DOI: 10.1016/j.atmosres.2024.107793 Shuailong Jiang, Yingying Ma, Fengdong Deng, Lianfa Lei
Atmospheric Research ( IF 4.5 ) Pub Date : 2024-11-17 , DOI: 10.1016/j.atmosres.2024.107793 Shuailong Jiang, Yingying Ma, Fengdong Deng, Lianfa Lei
Atmospheric temperature and humidity profiles are critical for understanding atmospheric complexity and environmental protection. To unify the inversion algorithms of different stations, this paper establishes optimal model parameters for various stations in China using the Back Propagation Neural Network (BPNN) model. Additionally, a novel BPNN-based Temperature and Humidity Profile Retrieval (BPNNTR) model is proposed to improve the accuracy of atmospheric temperature and humidity profile inversions. Firstly, sensitivity experiments using 31 stations alongside the MonoRTM and BPNN models, were conducted to identify the optimal model parameters. Subsequently, Microwave Radiometer (MWR) datasets from the Xi'an and Taiyuan stations were used to train the new BPNNTR model. The experimental results indicate that the optimal model parameters for atmospheric temperature profile retrieval include 2 hidden layers and 1024 nodes. The retrieval errors exhibit a Mean Absolute Error (MAE) of less than 1.9 °C, a Root Mean Square Error (RMSE) of less than 2.5 °C and a Pearson Correlation Coefficient (R) abrove 0.99. For humidity profile retrieval, the optimal parameters range from 1 to 4 hidden layers and 512 to 1024 nodes, with most regions achieving the best results with 2 or 3 hidden layers and 1024 nodes. In these cases, the MAE is below 12 %, the RMSE is below 16 %, and R exceeds 0.80. At the Xi'an station, the BPNNTR model improves temperature profile retrieval accuracy by 5 % under sunny days, while significantly improving humidity profile retrieval by 15 % on sunny days, 12.6 % on rainy days, and 12.5 % on cloudy days. At the Taiyuan station, the model boosts retrieval accuracy by over 15 % on rainy days and maintains stable performance for humidity profiles, improving fitting accuracy by more than 11 % under cloudy days. This research unifies inversion algorithms across different sites and significantly enhances the inversion accuracy, benefiting environmental protection and atmospheric research.
更新日期：2024-11-17
京公网安备 11010802027423号