Passive microwave remote sensing of Sea Surface Salinity (SSS) in coastal regions has long been affected by severe Radio Frequency Interference (RFI) contamination. Coupled with inadequate physical parameterization of surface roughness effects, traditional SSS inversion methods often perform unsatisfactory in coastal areas. To overcome these persistent limitations while fully exploiting the multi-angle observation capability of Interferometric Microwave Radiometer (IMR) instruments, this study proposes a novel physics-guided machine learning framework for robust and accurate SSS retrieval. The key component of our approach is the Fully Convolutional Network with Global Receptive Field (FCN-GRF) architecture. Unlike Artificial Neural Networks (ANNs) that requires fixed-size inputs, the FCN inherently accommodates inputs of arbitrary sizes, thereby preserving the complete set of multi-angle Brightness Temperature (BT) measurements for each geo-grid without resorting to suboptimal interpolation or discarding valuable angular information. A key architectural innovation is the deployment of global kernels that span the full angular range, ensuring a global receptive field (GRF) and allowing the model to integrate the entire BT profile effectively. Crucially, the model is guided by physical insights through a carefully designed input tensor that integrates RFIcorrected BT measurements, physically calculated flat-surface BT estimates, and geographical proxy variables representing surface roughness effects. Furthermore, by broadcasting scalar variables to match the angular dimension, the model learns their angle-dependent modulations on the observed signal, effectively embedding physical constraints into the learning process while maintaining a streamlined workflow. Evaluations conducted over the marginal Northwest Pacific demonstrate the exceptional performance of the proposed SSS retrieval framework. It achieves Mean Absolute Error (MAE) of 0.576 psu against Copernicus Marine Service (CMEMS) reanalysis data, meeting the SMOS (Soil Moisture and Ocean Salinity) specified single-observation accuracy target of 0.5–1.5 psu, which the current SMOS SSS product fails to reach in the study region. Validation against the Global Temperature and Salinity Profile Programme (GTSPP) insitu observations further confirms the high fidelity of our results, with mean bias and standard deviation even lower than those of CMEMS reanalysis data, and substantially better than those of SMOS products. Moreover, the model exhibits generalization ability when applied to geographically and environmentally distinct regions with acceptable performance degradation, highlighting its transferability and robustness.