Cloud Detection Fusion Algorithm for Complex and Variable Surface Conditions

Cloud detection methods for Landsat satellites have been developed to characterize surface backgrounds and various cloud characteristics, employing different processing techniques to improve the accuracy of cloud detection. While these algorithms have their own strengths and limitations, it is crucial to improve their accuracy and stability through algorithm fusion. To address the issue of low detection accuracy for highly reflective surfaces and thin broken clouds, this study examines the impact of complex surfaces and seasonal variations and proposes a cloud detection fusion algorithm that combines multiple algorithms. First, the accuracy feedback indicators for each algorithm were calculated by combining GlobeLand30 and Landsat 8 Cloud Cover Assessment Validation Data. A foundational fusion algorithm model was constructed based on the accuracy feedback indicators. Subsequently, considering the influence of complex surfaces and seasonal variations, the normalized vegetation index was introduced into the basic fusion algorithm. The algorithm weights were dynamically updated in real time using logistic regression. The detection accuracy of the fusion algorithm was evaluated using Landsat 8 Biome data and Landsat 9 images. The fusion algorithm exhibited a high consistency of 0.95 for real cloud coverage. Its F1 score was 0.96 for water and 0.92 for barren surfaces. The results demonstrate that the fusion algorithm can identify thin and broken clouds above both dark and bright surfaces with high accuracy. The fusion algorithm outperformed other similar algorithms, as it could overcome the influence of complex surfaces and seasonal variations.