Scenario-Adaptive Deep Joint Source-Channel Coding for CSI Feedback

Multiple-input multiple-output (MIMO) technology is expected to be a critical technology for next-generation wireless communication systems. A major challenge in frequency division duplexing (FDD) MIMO systems is the realization of low-overhead, high-spectral-efficiency channel state information (CSI) feedback. The advent of deep learning has introduced a promising paradigm for artificial intelligence (AI)-based CSI feedback enhancement and has drawn attention in the standardization of AI-native wireless communication systems. However, existing studies still face the critical challenge of limited generalizability across diverse channel scenarios. CSI feedback networks trained for specific environments often suffer from degraded performance under channel scenario changes, or require model switching, which introduces latency and results in increased model storage overhead during practical deployment. In this paper, we propose a scenario-adaptive deep joint source-channel coding (DJSCC) architecture. Channel scenario information is incorporated as auxiliary information to enable the network to adapt both feature extraction and feedback bandwidth according to the communication scenario. The proposed network architecture, named JSBANet, achieves simultaneous scenario-adaptation and bandwidth-adaptation, thereby improving CSI reconstruction accuracy while reducing feedback overhead. Extensive experiments on multi-scenario CSI datasets validate the effectiveness of the proposed method and demonstrate its performance advantages over traditional and baseline AI-based CSI feedback approaches.