Improved cloud detection for Landsat 8 images using a combined neural network model

High-precision cloud detection is a key step in the processing of remote sensing imagery. However, the existing cloud detection methods struggle to extract high-accuracy cloud pixels, especially for images of thin and fragmented clouds or those over high-brightness surfaces. In this study, we developed a new model by combining the existing models of Fully Convolutional Network-8 sample (FCN-8s) and U-network (U-net) (based on the three visible bands) to take full advantage of spectral and spatial information. In the proposed Fully Convolutional Network Ensembling Learning (FCNEL) model, U-net and FCN-8s initially conduct separate classifications based on their relative strengths, and their outputs are fused by the voting strategy to integrate multi-scale features from both the models. Different surface and cloud types in Landsat 8 Operational Land Imager (OLI) data were used to test the model, which showed an average overall accuracy of 91.68% and an average producer accuracy of 98.52%. Thus, the proposed FCNEL model was superior to FCN-8s or U-net as well as the widely used function of mask algorithm. The proposed method has good adaptability to various cloud types and diverse underlying surface environments.