Reconstructing spectral information in cloud contaminated regions using object-oriented approach

Optical remote sensing imagery can serve as a crucial data source in numerous fields. However, the susceptibility to weather conditions (like clouds and rain) can render the cloud contamination, leading to the significant influence on optical images. The image quality can be required to restore the spectral information in cloud-contaminated regions. In this study, the spectral information was reconstructed in the cloud-contaminated regions using an object-oriented approach. Synthetic aperture radar (SAR) data was integrated with the radar signals and land cover information, in order to serve as the constraints for the reconstruction. The distance matrix was calculated on the SAR images, and then transplant onto the optical images for the replacement. The pixel-level matching accuracy was enhanced to maximize the retention of spectral information from the original images. A solid foundation was provided for the accurate extraction of subsequent land surface information. The performance was also evaluated to verify the improved model. The comparative experiments were conducted with the traditional pixel replacement, weighted linear regression, and curvature-driven. Visual inspection was carried out over various scenes. The results revealed that the superior quality was achieved in the imagery reconstruction in the thick cloud areas and cloud shadows. Particularly, the consistency was maintained among various land cover pixels in the reconstructed areas by the object-oriented method. The boundaries and areas were accurately delineated in the various types of fine-grained land cover. Quantitative analysis was conducted to validate using some metrics, such as correlation coefficient, root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). The spectral characteristic curves of the pixels were reconstructed to closely match those of the reference images, with the correlation coefficients exceeding 0.99. Notably, significant improvement was observed in the spectral curve recovery of vegetation pixels with the enhanced relative to the rest. The radar vegetation indices were optimized to employ an object-oriented matching mechanism during pixel reconstruction. The traditional pixel replacement model was optimized effectively. Furthermore, the band reflectance was closely resembled that of the reference images. Among the five bands of cloud-contaminated images after reconstruction, the highest correlation coefficients were observed with Band 2 and Band 3 reaching 0.925 and 0.922, respectively, and RMSE values of 0.009 and 0.007, respectively. The reconstructed images shared the minimal quality loss and high similarity to the reference images, with the highest PSNR and SSIM values across all bands. Specifically, the structural similarity values of the reconstructed Band 2 and Band 3 images were 0.902 and 0.910, respectively, with the PSNR values reaching 31.978 and 33.173, respectively. The superior image quality was obtained similar to reference images, which was consistence with spectral characteristics. The experimental results were used to the qualitatively and quantitatively validated the effectiveness of the reconstruction. The object-oriented approach can be expected to reconstruct the spectral information in the cloud-contaminated regions. The remotely sensed images were consistently produced with the higher spectral reconstruction accuracy in the different types of land cover and spectral bands. However, the replacing pixel spectral information was relied primarily on from single-image data. Some limitations was matched the effectiveness between cloudy and cloud-free regions, when reconstructing the large-scale, thick cloud areas in single remote sensing images. Additionally, the global searching can fail to the optimal reconstruction for the locally contaminated cloud pixels. Therefore, further investigation is required to determine the optimal local search range for the pixel matching using different extents of cloud contamination.