PPG-Net: Polarization Physics-Guided Dehazing for Near-Surface Optical Remote Sensing
Abstract
Atmospheric haze attenuates target radiance and adds path radiance, obscuring land-surface structure in optical imagery. We present PPG-Net, a polarization physics-guided image-dehazing network that converts four analyzer-angle observations into Stokes components, polarization features, a coarse transmission map, and a coarse physical reconstruction. A multiscale stem and dilated residual backbone jointly predict scene radiance and transmission. Inference-time fusion combines the direct restoration, transmission-based inversion, and a dark-channel restoration. Evaluation uses synthetic polarimetric haze generated from four terrestrial RGB-D sources, with splits disjoint at the source RGB-D image level. This short-range benchmark does not reproduce airborne or satellite atmospheric paths. PPG-Net obtains 23.11 dB PSNR and 0.8290 SSIM on the test set. Budget-matched retraining of multiple control variants evaluates the contributions of input representation and transmission supervision before inference-time fusion. The results reveal a tradeoff between image fidelity and physical interpretability. Fusion improves the complete model’s direct output. Six public real polarimetric examples show clearer contours but retain darkening and color bias. These results support polarization-guided restoration under the tested synthetic conditions; deployment in near-surface optical remote sensing still requires calibrated real-data validation.