Abstract: For spaceborne synthetic aperture microwave array radiometers, the limitation in the number of antennas often results in Gibbs oscillations when retrieving brightness temperature images. This causes the sea-land boundaries in the images to become blurred, making it difficult to fully align with the real boundaries. In recent years, convolutional neural networks (CNN) have been widely applied in the field of remote sensing image processing, providing new ideas for solving this problem. However, the training datasets for traditional CNN are mostly generated based on optical images, which differ significantly from the sea-land brightness temperature images obtained by spaceborne synthetic aperture microwave radiometers. This inevitably affects the training effectiveness of the models. To address the aforementioned issues, this paper first proposes a method for generating simulated brightness temperature observation images through a forward model. This method, based on real sea-land information, can obtain a large number of highly realistic simulated brightness temperature images observed by synthetic aperture microwave radiometers. Additionally, the paper optimizes and improves the structure of the CNN model to effectively prevent overfitting. Compared to traditional windowing methods, the proposed nearshore brightness temperature error correction method based on the CNN model significantly improves image resolution and reduces Gibbs oscillation effects, resulting in a noticeable enhancement in image quality.