Lightweight Remote Sensing Multispectral Image Compression With Decoupled Interactive Knowledge Distillation
Qizhi Fang, Jingang Wang, Jiahui Liu, Lili Zhang · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
To address the challenges of spectral and spatial redundancy in multispectral images, as well as the limitations of satellite-to-ground bandwidth and onboard computational resources, a lightweight multispectral image compression with decoupled interactive knowledge distillation framework, MIC-DIKD, is proposed in this particleaper. Within the encoder–decoder backbone, a multiscale spatial–spectral redundancy reduction block and a decoupled dual-branch attention module are introduced. Cross-band and cross-spatial dependencies are jointly modeled. Redundancy is suppressed at the source, and representational capacity is enhanced. Dual constraints at the pixel and feature levels are further applied. The student model is guided toward the teacher. Strong rate–distortion performance and perceptual quality are achieved. Experiments on Landsat-8 and Sentinel-2 show that MIC-DIKD outperforms several baselines in peak signal-to-noise ratio, multiscale structural similarity, and mean spectral angle. In terms of complexity, the student model distilled from the teacher provides comparable performance. FLOPs and encoding/decoding latency are substantially reduced. The framework is therefore well suited to the engineering requirements of onboard compression and ground-based fast decoding. Overall, a practical solution is provided for multispectral image onboard compression, ground decoding, and large-scale processing. High performance is maintained, and complexity is reduced.