Shadow-Effect Correction in Aerial Color Imagery

Due to the existence of shadows, especially in urban environments, it is difficult to extract semantic information from aerial and high-resolution satellite images. In this paper, an efficient method of correcting shadow effects using multisource data sets in aerial color images is proposed. The proposed method has three steps. First, it accurately detects the shadowed regions using the image geometry and the solar position of the image acquisition data. Then, the detected shadowed regions are segmented according to land surface type. Finally, the shadow effects of the segmented regions are corrected by directly comparing the same nonshadow features with the segmented shadows. In the application part of this paper, the proposed techniques were applied in the extraction of an asphalt road from an image. Introduction In remotely sensed imagery, shadows can sometimes be used to extract geospatial information. However, shadows sometimes have a negative effect on the extraction of geometric and semantic information on the earth’s surface especially in urban environments, because shadows cast by high-rise ground objects make it difficult to implement photogrammetric applications such as feature extraction, automatic triangulation, and orthophoto generation. Intensive research has been implemented to correct shadow effects in remotely sensed images (Pouch and Compagna, 1990; Colby, 1991; Itten and Meyer, 1993; Richter, 1998; Kawata et al., 1998; Simpson and Stitt, 1998; Rau et al., 2002; Dare, 2005). Most research on shadow detection and removal has been done, however, on mountainous terrain using satellite images (Hall-Konyves, 1987; Civco, 1989; Lui and Moore, 1993; Wang et al., 1999; Giles, 2001). Because mountainous areas are composed mainly of forests, the correction of shadow effects is relatively easy in this regard compared to that on images with complex built-up urban environments. Simpson and Stitt (1998) removed cloud shadows from AVHRR data using geometric considerations to project the cloud features on the ground in the direction of a solar azimuth angle and the cloud height. Wang et al. (1999) proposed the automated algorithm to detect and remove cloud shadows from Landsat TM data based on changes in the reflectance and frequency Shadow-Effect Correction in Aerial Color Imagery Hong-Gyoo Sohn and Kong-Hyun Yun components and replacement with the acquired non-shadow image at a different time. Giles (2001) also investigated the shadow detection algorithm from a Landsat TM image and described a quantitative spatial evaluation of the proposed method in comparison to manual interpretation. On the treatment of shadows in urban areas, Rau et al. (2002) have observed that shadow effects could be corrected using the histogram matching method, which is applied to minimize the gray value differences between a shadowed area and its surroundings in the process of producing true orthoimages. The histogram matching method is also used to remove shadows in aerial photographs (Shu and Freeman, 1990). Moreover, Dare (2005) showed that shadows could be removed with the histogram adjustment method in built-up urban areas. Although, the results can be image-dependent. In addition to the methods used in the abovementioned researches, novel methods such as spectral end-members and the matched filter concept are used to remove shadows (Boardman, 1993; Adler-Golden et al., 2002). This study attempts to correct shadow effects in aerial color imagery using multi-source data sets. To detect the shadowed regions, digital maps and lidar data were selected to generate Digital Elevation Model (DEM) and Digital Surface Model (DSM), respectively. To avoid confusion of acronyms such as DEM and DSM in this paper, DEM was defined as a model that represents a topographic surface, and DSM, as a model that delineates the canopy of the object surface. The shadowed regions were accurately modeled using the geometric relationship between a three-dimensional model of the research area and the solar position of the image acquisition data. Also, to accurately correct the shadow effects, the shadowed regions were segmented using an existing digital map, which included the land surface cover types. Afterwards, based on the three proposed basic assumptions, the identified shadowed regions were given new calculated values. In the last part of this paper, to evaluate the proposed scheme, the potential application of asphalt road extraction using only a digital aerial image and lidar data is presented. Data Preparation The study site and the data sets were carefully selected for this research to test the proposed method. The used data sets consisted of a mosaic of aerial color imagery, lidar data, and digital maps with a scale of 1:1000. The study site was in some parts of Sungnam City (approximately 127.01°E, 37.36°N), Korea. The aerial color image was acquired on PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING May 2008 611 Hong-Gyoo Sohn is with the School of Civil and Environmental Engineering, Yonsei University, 134 Shinchon-Dong Seodaemun-Gu, Seoul,120-749, Korea. Kong-Hyun Yun is with the Construction Engineering Research Institute, Yonsei University, 134 Shinchon-Dong Seodaemun-Gu, Seoul,120-749, Korea ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 74, No. 5, May 2008, pp. 611–618. 0099-1112/08/7405–0611/$3.00/0 © 2008 American Society for Photogrammetry and Remote Sensing 05-064.qxd 4/11/08 3:17 PM Page 611