Applying Shadow Removal Technique for Urban Area Identification of High-Resolution Aerial/Satellite Images Using Color Information and Deep Learning

Urban mapping using high-resolution aerial and satellite images presents numerous opportunities. However, in high-density urban areas, crucial information is often obscured by shadows cast by tall buildings, leading to compromised classification results or misinterpretations and partial or complete loss of information within shadowed regions. While shadow removal in remote sensing data has been extensively studied, few studies have specifically addressed these issues in high-resolution satellite images. This research proposes an automated system for shadow identification and removal to mitigate the impact of shadows and enhance the differentiation of urban targets in multispectral images. The image bands are pre-processed to enhance information, followed by a shadow detection process using morphological filtering and Python coding for deep learning to generate a shadow mask. This process is accelerated using GPU programming. The images are then converted from RGB to the CIELCh model to create shadow masks, with a threshold value determined using the K-Means clustering algorithm. The utilization of these shadow masks assists in eliminating shadows from the original images by considering the lighting balance between shadowed and non-shadowed areas. This process enables the restoration of parts obscured by shadows. The experiments were conducted on the ISPRS Toronto and WHU building datasets, yielding highly effective results.