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

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cloud detection algorithms in remote sensing utilize spectral and spatial data.
• Existing methods have limitations like misclassifying bright surfaces and omitting thin clouds, leading to inconsistent performance.
• A need exists to integrate multiple deep learning models to overcome individual weaknesses.

Purpose of the Study:

• To propose a novel ensemble framework for improved cloud detection.
• To systematically integrate complementary strengths of multiple deep learning models.
• To address the research gap in robust, large-scale cloud detection.

Main Methods:

• Evaluated three deep learning algorithms (CD-SLCNN, SRMF-CD, CNN-TransNet) on Landsat-8 imagery.
• Developed an F1-score-weighted fusion strategy using algorithm-specific weights from a validation set.
• Created a weighted probabilistic fusion for enhanced cloud detection.

Main Results:

• The proposed F1-weighted ensemble outperformed standard ensemble baselines.
• Achieved a 4.7% point improvement in F1-score over majority voting.
• Demonstrated a 1.9% point improvement over Bayesian model averaging, reducing errors in heterogeneous landscapes.

Conclusions:

• The F1-weighted ensemble framework offers a robust and accurate solution for large-scale cloud detection.
• This approach effectively mitigates individual model weaknesses by leveraging their complementary strengths.
• The method enhances performance across heterogeneous landscapes, improving overall reliability.