Image-Domain Deep-Learning LSRTM With Multiscale Degradation Modulation and Forward Reconstruction Consistency

Reverse time migration (RTM) images are jointly constrained by limited acquisition aperture, band-limited source wavelets, and errors in the background velocity model. These factors often cause spatially varying resolution loss, uneven illumination, and amplitude distortion. These problems are more obvious in deep structures and poorly illuminated areas. Data-domain least-squares RTM (LSRTM) improves imaging resolution effectively. However, repeated migration and demigration lead to high computational cost and low efficiency. To address this problem, this article proposes an image-domain learning-based LSRTM (DL-ID-LSRTM) framework. The framework introduces degradation information related to migration conditions as conditional constraints during inversion. A forward reconstruction consistency (FRC) constraint is also introduced to improve image-domain consistency. In particular, the proposed method fuses the migration velocity model with a point spread function (PSF) descriptor. These inputs parameterize local degradation features caused by migration. These features depict local focusing behavior and illumination patterns. A spatially adaptive feature modulation (SAFM) mechanism then injects these conditional features into a multiscale inversion backbone. This design guides reflectivity reconstruction. In addition, this article designs a closed-loop forward reconstruction branch. Under the same migration conditions, the predicted reflectivity is used to synthesize an RTM profile, and an image-domain consistency loss is imposed between this synthesized profile and the input RTM image. This design suppresses shortcut learning and reduces nonphysical artifacts. Results from benchmark tests on two synthetic datasets, one out-of-distribution generalization experiment, and one marine field dataset support the proposed method. Compared with RTM, U-Net-based image-domain LSRTM, and data-domain LSRTM with a limited number of iterations, DL-ID-LSRTM produces estimates closer to subsurface reflectivity features. The results show improved lateral reflector continuity, restored amplitudes in weakly illuminated deep zones, reduced migration-related smearing, and a more balanced relative amplitude distribution with depth.