AFAR-Net: Autoregressive and feedback-driven adaptive rectangular convolution network for pansharpening

To enhance spectral fidelity and spatial detail restoration in remote sensing pansharpening tasks, this paper proposes a deep pansharpening network based on an encoder-decoder architecture, named the Autoregressive and Feedback-Driven Adaptive Rectangular Convolution Network for Pansharpening (AFAR-Net). The proposed network employs an autoregressive mechanism, where the output of the previous unit is used to optimize the current one, enabling progressive multi-level image reconstruction. Meanwhile, a feedback-driven fusion module is designed to efficiently integrate deep features across units, thereby enhancing spectral consistency. On the other hand, an adaptive convolutional residual block is introduced to flexibly adjust kernel sizes and shapes, strengthening the network's ability to model and restore complex spatial structures. Finally, a lightweight fusion head is utilized to aggregate multi-level predictions, improving the stability of reconstruction. Experimental results on multiple remote sensing datasets demonstrate that the proposed network outperforms existing mainstream approaches in terms of spatial distortion, Spectral Angle Mapper (SAM), and the Quality with No Reference (QNR) index, showing strong generalization ability and application potential.