A physics-constrained superpixel-based multi-source remote sensing spatiotemporal fusion network

Remote sensing spatiotemporal fusion aims to generate image sequences with both high spatial and temporal resolution. However, many deep learning approaches treat fusion as a black-box regression problem, overlooking differences in sensor imaging physics—such as the point spread function (PSF) and pixel-area integration—which often leads to boundary blurring, spectral mixing, or even physically implausible high-frequency details over heterogeneous landscapes. In addition, these methods commonly rely on paired coarse-/fine-resolution reference images at the same time, which limits data availability. To address these issues, we propose SPP-STF, a superpixel-based fusion network trained with a physics-guided degradation-consistency regularizer that requires only a reference fine-resolution image and a target coarse-resolution observation, while deriving an edge intensity map as a structural prior. The model adopts a U-Net-style, three-branch encoder–decoder to fuse texture, temporal variation, and edge cues at multiple scales. An EdgeResBlock uses high-frequency gating to sharpen boundaries and reduce cross-class leakage, and a superpixel cross-attention Transformer learns object-level correspondences to alleviate alignment errors caused by mixed pixels. During training, we introduce a learnable anisotropic Gaussian PSF-based degradation-consistency term: the predicted fine-resolution output is blurred and downsampled via area integration to match the observed coarse image, and this term is optimized jointly with the reconstruction loss. Quantitatively, SPP-STF achieves the best RMSE, SAM, PSNR, and UIQI on CIA (0.0254, 0.1522, 31.9330 dB, and 0.8451), the best results across all six metrics on LGC (including RMSE 0.0188 and PSNR 34.6307 dB), and the best RMSE, SSIM, SAM, PSNR, and ERGAS on Tianjin (0.0410, 0.7640, 0.2459, 27.8666 dB, and 30.4838). Ablation studies confirm the complementary benefits of the superpixel and edge modules. The PSF term is interpreted here as a training-time regularizer that provides consistent and physically more reasonable improvements across the three datasets. Structured pruning further indicates that the model can retain spatial structure under substantial compression, while spectral fidelity becomes notably more sensitive beyond roughly 70% pruning.