A Coarse‐to‐fine Approach to Detect Shadows in the Chang’E‐4 VNIS Hyperspectral Images

The low value of shadow pixels in the images sensed by the Visible and Near-infrared Imaging Spectrometers (VNIS) aboard Chang'E-4 Yutu-2 rover leads to an unrealistic representation of lunar surface reflectance. Due to the shadow features and lack of additional information (e.g., topography and shadow labels), it is challenging to apply existing methods to detect shadows for VNIS data. Therefore, it is necessary to propose an effective way to detect and correct the shadows in the VNIS data. In this paper, a coarse-to-fine approach to detect shadows in the VNIS images is developed. First, to better resolve the shadows, they are enhanced based on principal component analysis and band combination. Second, in the coarse detection stage, two shadow indices, SId and SIs, based on the features extracted in shadow enhancement are constructed. Subsequently, they are segmented into shadow and nonshadow pixels individually using the Otsu method. Third, in the fine detection stage, by analyzing the performance of the two indices, the shadow and nonshadow pixels are further distinguished through the absolute z-score to obtain the final results. The experimental results show that the average overall accuracy is 93.94%. The average F1-score is 85.51%. By comparing with visual inspection and other detection methods, our approach yields a good precision and is expected to be applicable for other VNIS hyperspectral data. After detection, the shadow effect is corrected in the spectral dimension by estimating the loss of radiometric information. Our study can provide a reference for shadow detection of analogous lunar in-situ observations.