Retrieving The Hidden Details Under The Shadows On Satellite Images Using Machine Learning Technology

Abstract

Shadow detection and recovering hidden objects under the shadows is one of the important techniques needed to increase the clarity of highly discriminating satellite images. The currently used techniques to remove shadows suffer from the edge effect of the detected shadows, and the use of training limited to a set of data leads to color inconsistency in the shadow area. In this research, a new method structure was proposed based on the diving pyramid network in which a series of clustering operations are performed, and image features are grouped hierarchically for monitoring and prediction respectively. SRD Standard Reference Data was used that contains one and a half million images (devoid of Shadow / shadow images). Results were evaluated using image enhancement factors, such as Structural Similarity Index Measure In Shadow Area (SSIM-S) and PSNR The Peak-Signal-To-Noise In Shadow Area and when using the Dive and Neighbor Pyramid Network. The closest proposed method yielded the best results especially for SSIM-S and  PSNR- S values.