Study on X-band Dual Polarization Radar Echo Attenuation Correction Based on Transformer Architecture

Attenuation effects in X-band weather radar significantly constrain its detection accuracy.Traditional attenuation correction methods typically rely on empirical formulas with limited parameters and poor generalization capability， leading to significant uncertainty in the correction results.In recent years， deep learning algorithm with powerful nonlinear fitting capacity has emerged as a promising technical approach to overcome the limitations of conventional methods.Based on the Transformer's underlying framework， this study develops an X-band radar attenuation correction architecture named as XCORnet.Utilizing the observational data from the upgraded polarimetric S-band Next Generation Weather Radar （CINRAD/SAD） at Beijing Daxing in 2023 -2024 flood season as truth， the horizontal reflectivity （Z H） and differential reflectivity （Z DR） measurements in corresponding range bins from the Beijing Fangshan X-band dual-polarization radar （XPOL） are spatially and temporally matched.These matched XPOL observations of Z H， Z DR along with specific differential phase （K DP） are incorporated to construct the AI training dataset， in which 2642624 samples for Z H and 2605583 samples for Z DR， respectively.The dataset is partitioned with 80% for training and 20% for testing.Within the XCORnet framework， the models with K DP as the primary feature input for Z H and Z DR attenuation correction are trained， respectively.And then， two XCORnet-based models are evaluated by means of the test set.The results demonstrate that the AI-based model outperforms traditional methods significantly.For Z H correction， the ratio bias （BIAS） of XPOL to SAD are from original 0.875 to model-based correction 0.972， surpassing the empirical formula-based 0.901.The root mean square error （RMSE） are reduced from original 8.693 to model-based 5.811 dB with a 33.15% improvement， whereas the empirical formula only reduced it to 6.820 dB （with a 21.54% improvement）.For Z DR correction， the BIAS are from original 0.862 to model-based 1.141， outperforming the over-correction of empirical formula-based （1.273）.The RMSE decreased from original 1.679 to model-based 0.972 dB （with a 42.10% improvement）， compared to the empirical formula-based reduction only 1.382 dB （with a 17.69% improvement）.For Z H correction， the mean absolute error （MAE） improves from 6.292 to 4.222 dB （with a 32.89% improvement）， whereas the empirical formula only reduces it to 5.113 dBZ （with a 18.73% improvement）.For Z DR correction， the MAE decreases from original 1.271 to model-based 0.697 dB （with a 45.16% improvement）， compared to the empirical formula-based reduction only to 1.008 dB （with a 20.69% improvement）.The MAE with model correction also showed distinct advantages over that with traditional methods.Three cases further validate the stability and generalization capability of AI-based correction.