Fast spatial-spectral random forests for thick cloud removal of hyperspectral images

The long-term existence and extensive coverage of thick clouds have caused an amount of missing information in hyperspectral images (HSIs), hindering their applications greatly. Recently, with the continuous launch of new satellites, the increasing availability of HSIs provides more opportunities for using temporal information in cloud removal of HSIs. In this paper, a fast spatial-spectral random forests (FSSRF) method was proposed for removing clouds of HSIs. FSSRF is developed based on the advanced spatial-spectral random forests (SSRF) method that was developed to handle multispectral images. FSSRF greatly improves the computing efficiency while ensuring the accuracy of reconstruction. Experimental results on both GF-5 and EO-1 HSIs show that FSSRF is a much faster version than the original SSRF that uses directly all known bands of the temporally neighboring HSI with spatially complete coverage, and the accuracies of the two versions are similar. Compared with subSSRF (i.e., SSRF using several spectrally adjacent known bands for each cloudy band) and the popular MNSPI method, FSSRF can produce more accurate predictions. The evaluation from spectral dimension shows that FSSRF can recover the spectral characteristics of cloud pixels more satisfactorily. FSSRF has great potential in real-time cloud removal of HSIs due to its obvious advantages in balancing computational efficiency and accuracy.