CLEAR: A Fast Cloud Removal Method for Optical Remote Sensing Images Combining Class-based Linear Regression and Iterative Residual Compensation

Cloud and cloud shadow contamination significantly reduce the applicability of optical remote sensing images, introducing relevant gaps in land-cover representation. Existing cloud removal methods usually do not achieve both high accuracy and high computational efficiency, limiting their effectiveness in real applications. In this article, we propose a fast and effective method named class-based linear regression and iterative residual compensation (CLEAR) to fill gaps in optical images. In the class-based linear regression step, CLEAR first classifies the stacked time-series reference images to effectively characterize the land-cover dynamics. It then preliminarily fills gaps by establishing the relationship between the reference images and the target cloudy image. The iterative residual compensation step further improves the cloud removal accuracy for each target cloudy pixel by integrating the residuals of its neighboring similar pixels. The similar pixel selection is achieved by a novel K -dimensional (K-D) tree-based approach. The performance of CLEAR was compared with seven representative cloud removal methods using simulated cloudy images with different cloud covers at three study sites. The experimental results demonstrate that CLEAR outperformed other methods by obtaining the highest prediction accuracies and preserving better spatial details in the cloud-removed images. Moreover, CLEAR also presented higher computational efficiency compared with the other methods. We also demonstrate CLEAR’s potential in supporting agricultural and large-scale remote sensing applications. The Python code of CLEAR and the experimental dataset are available at https://github.com/HoucaiGuo/CLEAR-code.

Cloud and cloud shadow contamination significantly reduce the applicability of optical remote sensing images, introducing relevant gaps in land-cover representation. Existing cloud removal methods usually do not achieve both high accuracy and high computational efficiency, limiting their effectiveness in real applications. In this article, we propose a fast and effective method named class-based linear regression and iterative residual compensation (CLEAR) to fill gaps in optical images. In the class-based linear regression step, CLEAR first classifies the stacked time-series reference images to effectively characterize the land-cover dynamics. It then preliminarily fills gaps by establishing the relationship between the reference images and the target cloudy image. The iterative residual compensation step further improves the cloud removal accuracy for each target cloudy pixel by integrating the residuals of its neighboring similar pixels. The similar pixel selection is achieved by a novel $K$ -dimensional (K-D) tree-based approach. The performance of CLEAR was compared with seven representative cloud removal methods using simulated cloudy images with different cloud covers at three study sites. The experimental results demonstrate that CLEAR outperformed other methods by obtaining the highest prediction accuracies and preserving better spatial details in the cloud-removed images. Moreover, CLEAR also presented higher computational efficiency compared with the other methods. We also demonstrate CLEAR's potential in supporting agricultural and large-scale remote sensing applications.