Abstract:

Suppressing water glint pollution from remote sensing images and reconstructing image information are effective ways to improve the quality of UAV (unmanned aerial vehicle) remote sensing information and increase water environment monitoring areas. It is difficult to apply traditional glint information reconstruction algorithms to UAV hyperspectral images. This study proposes an algorithm for automatic glint detection, removal, and information reconstruction. First, NDWI (normalized difference water index) was used to extract the water body, and the lowest value of the sum of grayscale images in the entire band was used as a threshold to segment the glint, and the Laplace operator was used to extract the glint texture. The difference between the two areas was calculated through multiple rounds of morphological expansion and threshold updates. The lowest difference occurrence frequency was obtained by voting, and the best threshold was obtained in reverse to remove the glint automatically. Then, we determined the matching bands based on principal component analysis and compared the minimum similarity of matching blocks of different sizes to obtain the best size of the image blocks. Finally, we used an improved Criminisi algorithm to reconstruct the flare removal region. The removal algorithm was applied to four real glint scenarios with a removal rate > 99%; the reconstruction algorithm results are superior to those of other algorithms both subjectively and objectively, and the difference between the variation coefficient of each band of the glint reconstruction for water and normal water was within 1%, indicating good spectral application capability.

Key words: UAV hyperspectral, glint removal, information reconstruction, principal component analysis, improvement of the Criminisi algorithm