Uncertainty-Aware Hybrid Ensemble Framework for Haze Degree Estimation in Vision-Based Processing Systems

Abstract Haze severely degrades outdoor image visibility and color fidelity, limiting the reliability of vision-based systems used in autonomous navigation, environmental monitoring, and aerial surveillance. Existing haze degree estimation approaches either rely on handcrafted priors with limited generalization or deep learning models that behave as black boxes without awareness of uncertainty. To address this gap, we present a Hybrid Haze Degree Estimation (HHDE) framework that integrates interpretable physical haze descriptors with high-level convolutional representations within a probabilistic ensemble learning paradigm. Nineteen haze-relevant features are fused with CNN embeddings and processed using a stacking architecture consisting of Random Forest and XGBoost regressors. At the same time, a Bayesian Ridge meta-learner produces uncertainty-aware haze predictions. HHDE quantifies both haze degree and its corresponding confidence interval, improving reliability in uncertain atmospheric conditions. Extensive evaluation on ten benchmark datasets demonstrates the superior performance of HHDE over classical regression algorithms and state-of-the-art deep learning models, achieving an MAE of 0.0206, an RMSE of 0.0301, and an $R^2$ of 0.9525. Ablation analyses confirm the complementary roles of hybrid feature fusion and Bayesian reasoning. The results establish HHDE as a robust, interpretable, and uncertainty-aware solution suitable for real-world environmental image analysis and autonomous vision systems.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-8370337/v1
Primary Topic
Image Enhancement Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Uncertainty-Aware Hybrid Ensemble Framework for Haze Degree Estimation in Vision-Based Processing Systems

Garima Kadian, Rajiv Kumar
Research Square
Image Enhancement Techniques
preprint

Uncertainty-Aware Hybrid Ensemble Framework for Haze Degree Estimation in Vision-Based Processing Systems

Garima Kadian, Rajiv Kumar
preprint en

Abstract

Abstract Haze severely degrades outdoor image visibility and color fidelity, limiting the reliability of vision-based systems used in autonomous navigation, environmental monitoring, and aerial surveillance. Existing haze degree estimation approaches either rely on handcrafted priors with limited generalization or deep learning models that behave as black boxes without awareness of uncertainty. To address this gap, we present a Hybrid Haze Degree Estimation (HHDE) framework that integrates interpretable physical haze descriptors with high-level convolutional representations within a probabilistic ensemble learning paradigm. Nineteen haze-relevant features are fused with CNN embeddings and processed using a stacking architecture consisting of Random Forest and XGBoost regressors. At the same time, a Bayesian Ridge meta-learner produces uncertainty-aware haze predictions. HHDE quantifies both haze degree and its corresponding confidence interval, improving reliability in uncertain atmospheric conditions. Extensive evaluation on ten benchmark datasets demonstrates the superior performance of HHDE over classical regression algorithms and state-of-the-art deep learning models, achieving an MAE of 0.0206, an RMSE of 0.0301, and an $R^2$ of 0.9525. Ablation analyses confirm the complementary roles of hybrid feature fusion and Bayesian reasoning. The results establish HHDE as a robust, interpretable, and uncertainty-aware solution suitable for real-world environmental image analysis and autonomous vision systems.

Research Square
Termez Institute of Engineering and Technology (UZ)
Image Enhancement Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon