Yang, Xu, Zuo, Xiaoqing, Zhu, Daming, Xie, Wenbin, Li, Yongfa, and Guo, Shipeng
Shadow indexes can effectively and easily detect shadows in remote sensing images. Although various shadow indexes are used, there is no consensus regarding the optimal one. The evaluation criteria used to determine the performance of shadow indexes are relatively homogeneous. A multi-criteria evaluation of widely used shadow indexes is necessary to analyse the application performance of each index in different scenarios. The eight most common shadow indexes were selected for this study. In addition to the standard visual inspection and accuracy assessment criteria, the cosine of solar incidence (cosi) was introduced to evaluate the sensitivity of the indexes to illumination intensity. Furthermore, in order to evaluate the separability of shadows and confusable objects, the interquartile range (IQR) of shadows and confusable objects was used to calculate the 1.5IQR separation ratio (1.5IQRSR). The multi-scale evaluation included mountain and urban scenarios, with the sensitivity to illumination intensity used to assess mountain scenarios and the capacity to separate confusable objects used to assess urban scenarios. The results obtained for each evaluation criterion were aggregated to form the multi-criteria evaluation score of the shadow indexes. The results show that the combinational shadow index (CSI), shadow detector index (SDI), and shadow index (SI) perform relatively well for each evaluation criterion. SI and CSI had the highest shadow detection accuracy in mountain and urban areas, respectively. The shadow enhancement index (SEI) and CSI had relatively high sensitivity to illumination intensity, especially CSI. In addition, CSI has the highest water separability but low separability for blue and dark impervious surfaces. SDI and SI have high separability for confusable objects, but they are ineffective in separating water bodies and less sensitive to illumination intensity. In summary, CSI, SDI, and SI excel in many application contexts, whereas the performances of the remaining five indexes were similar in all respects. [ABSTRACT FROM AUTHOR]