Abstract: In response to the problem that the sensor-based methods for detecting moisture distribution in grain bulks lack representative and fail to truly reflect internal moisture distribution, this study proposes a very high-frequency electromagnetic wave-based method for moisture distribution detection and three-dimensignal inversion in grain bulks.Based on a calculation model for the spatial distribution of dielectric constant is established based on the dense layering of electromagnetic wave echoes, and the moisture distribution inversion is realized by integrating the dielectric-moisture model. During model optimization, a depth adaptive correction factor is introduced to modify the recursive process, and a depth-aware noise weighting mechanism is proposed to enhance the signal-to-noise ratio in the deep layers. A multi-frequency electromagnetic wave combined detection system is constructed, and an attention-mechanism-weighted frequency feature fusion method is adopted to balance shallow-layer resolution and deep-layer penetration capability. Furthermore, an improved U-Net network was introduced to improve the recognition accuracy of abnormal moisture areas. The results show that within the depth range of 0.15-0.90 m, the standard deviation of the calculated dielectric constant is 0.09-0.11, and the root mean square error is 0.21-0.23. In the depth range exceeding 1 m, the dielectric constant calculation also achieves satisfactory results, and the constructed deep learning model can accurately segment abnormal moisture areas inside the grain bulk. This method effectively improves the accuracy and detection depth of grain bulk moisture distribution detection, providing strong technical support for grain storage safety.