Multi-Temporal Dehazing of Hyperspectral Remote Sensing Images

Haze removal from hyperspectral remote sensing imagery is an important restoration task. According to the atmospheric scattering model, haze attenuates the surface-reflected signal and introduces an atmospheric component, reducing image contrast and distorting spectral signatures. Existing hyperspectral dehazing methods rely on a single hazy observation, which limits recovery under dense haze when surface information is strongly degraded.

We investigate a multi-temporal dehazing approach where an auxiliary hyperspectral image of the same area, acquired at a different time, supports restoration of the target hazy image. The approach is implemented by modifying the single-image AACNet architecture to incorporate auxiliary information. We also construct a multi-temporal hyperspectral dataset containing synthetic haze for controlled evaluation and real haze for assessing performance under natural temporal differences.

Restoration quality was evaluated using PSNR, SSIM, and SAM. On the synthetic test set, single-image AACNet achieved 34.226 dB/0.818/0.116. The proposed model achieved 40.353 dB/0.857/0.059 with temporally more distant auxiliary images and 45.833 dB/0.867 / 0.041 with temporally closer images. On the real test set, it achieved 26.877 dB/0.667/0.202, compared with 26.048 dB/0.662/0.218 for single-image AACNet. The smaller improvement on real data likely reflects both the synthetic-to-real haze gap and scene changes between acquisitions; moreover, real hazy images and clean references are themselves acquired at different times, which can increase full-reference error.

At equal network depth, the multi-temporal model increases parameter count and computational complexity by only 0.41% and 0.44%, respectively, with an estimated memory increase of approximately 0.7%.

These results show that multi-temporal information can substantially improve hyperspectral dehazing, while its benefit depends on temporal consistency between observations.