Abstract: Objectives In recent years, X-band marine radar has been widely used in the study of ocean surface wave time-series extraction. However, both traditional spectral methods and the more recent deep learning approaches still exhibit significant limitations in terms of wave field reconstruction accuracy and parameter generalization. Moreover, most existing studies, both domestic and international, focus primarily on linear wave fields and only consider shadowing and tilt modulation effects in radar imagery, which fail to accurately capture the nonlinear characteristics of real ocean waves and corresponding radar backscattering behavior. Methods To address these problems, this study proposes an improved method that integrates fully nonlinear wave theory with deep learning techniques to achieve high-precision sea surface reconstruction from X-band radar images of wave fields. The proposed approach innovatively employs Bidimensional Empirical Mode Decomposition (BEMD) to decompose radar images in the time domain, extracting a set of Intrinsic Mode Functions (IMFs) that represent features at different frequency scales, along with residual component. These components are then fed into a U-Net architecture enhanced with both spatial and channel/t/nattention mechanisms. Finally, a series of experiments are conducted to systematically compare the performance of the improved method with baseline models from multiple perspectives. Results The results demonstrate that the improved method exhibits superior reconstruction performance and more robust generalization compared to the baseline models under complex sea conditions.Conclusions This validates the promising potential of our proposed improved method in the field of X-band radar-based wave field inversion.