An F1-score-weighted ensemble of deep learning models for enhanced cloud detection in remote sensing imagery

Cloud detection algorithms in remote sensing imagery exploit spectral and spatial characteristics, either independently or synergistically, to facilitate the discrimination between cloud and underlying surfaces. Individual cloud detection methods manifest unique operational advantages but suffer from inherent limitations, such as bright surface misclassification and thin cloud omission, leading to unstable performance across heterogeneous landscapes. A critical research gap exists in how to systematically integrate complementary strengths of multiple deep learning models while mitigating their individual weaknesses. To address this gap, this paper proposes a novel ensemble framework based on multiple deep learning models, using an F1-score-weighted fusion strategy. The framework first evaluates three representative deep learning-based detection algorithms (CD-SLCNN, SRMF-CD, CNN-TransNet) on Landsat-8 imagery, then derives algorithm-specific weights from their F1-scores on a validation set to produce a weighted probabilistic fusion. Quantitative comparisons on a hold-out test set show that the proposed method outperforms standard ensemble baselines: it improves F1-score by 4.7% points over majority voting and by 1.9% points over bayesian model averaging, while also reducing misclassification and omission errors across heterogeneous landscapes. The proposed F1-weighted ensemble thus provides a robust and accurate solution for large-scale cloud detection.