A Deep Learning Approach for Retrieving Cloud Microphysical Parameters with Multi-Frequency Cloud Radar

High-precision retrieval of cloud microphysical parameters is crucial for advancing cloud physics and improving weather and climate models. To address observational data scarcity, this study develops a multi-frequency residual attention network (MiRA-Net) based on forward simulations from X/Ka/W-band cloud radars. The model retrieves liquid water content (LWC), ice water content (IWC), effective radius (Re), and particle type (Ptype), and is evaluated against polynomial fitting (polyfit) and random forest (RF) methods. Results demonstrate that multi-frequency observations significantly enhance retrieval accuracy. MiRA-Net outperforms the others, achieving an RMSE of 0.178 for LWC, substantially lower than RF (0.251) and polyfit (0.256). With dual-frequency (W/Ka) linear depolarization ratio (LDR), MiRA-Net's IWC retrieval RMSE reaches 0.038—a 63% improvement over RF—with correlation near 1.000. Even without LDR, the RMSE remains as low as 0.054, indicating excellent robustness. Validation using independent field observations confirms its practical value, with IWC retrieval correlations of 0.78 and 0.87 in two cases.