Research on cloud detection method based on voting ensemble learning using FY-4B satellite data

Machine learning is widely used for satellite cloud detection. Ensemble models generally offer superior generalization over single models in complex scenarios. Addressing the common use of homogeneous ensembles, this study introduces heterogeneous base learners (SVM, NB, LR, DT, RF, MLP) trained on cloud features from FY-4B AGRI infrared data. We propose a Voting Ensemble Learning method to enhance generalization and adaptability for stable cloud detection. Results show that while RF and MLP excel individually, the Voting Ensemble significantly boosts overall accuracy and stability. Cross-validation with MODIS data confirms >91% accuracy for the ensemble over deep sea, shallow water, land, and snow, with false alarm rates generally <8% (except 12% for snow). The ensemble performs stably across seasons. Compared to single models, the Voting Ensemble demonstrates markedly improved adaptability, accuracy, and robustness across diverse seasons and scenarios, offering a more reliable solution for complex cloud detection challenges.