Conception et développement d’indicateurs intelligents de qualité de l’air pour soutenir la préservation de l’environnement et de la santé

Air pollution poses an escalating global threat to public health and environmental sustainability, with traditional monitoring systems struggling to provide the timely, accurate, and interpretable information essential for effective decision-making. This thesis addresses these critical limitations by developing intelligent air quality indicators that integrate advanced artificial intelligence techniques with environmental science to create more effective tools for environmental preservation and health protection.This research makes three fundamental contributions to environmental intelligence systems. First, we introduce a novel granular computing framework for explainable air quality classification that transforms opaque machine learning models into transparent decision support systems. We propose the AirQ-RuleGrCEx algorithm as our primary theoretical contribution a groundbreaking approach that systematically extracts human-readable rules by leveraging granular computing principles to explicitly link meteorological conditions to air quality levels. This algorithm represents a significant methodological advancement in the field of explainable AI for environmental applications, achieving competitive accuracy while providing unprecedented interpretability for environmental stakeholders and policymakers. Through systematic development and refinement, we aim to establish this approach as one of the most promising algorithms for interpretable environmental classification, bridging the critical gap between predictive performance and decision transparency.Second, we develop a comprehensive neural transfer learning paradigm through three innovative applications that revolutionize air quality forecasting efficiency and scalability. Our first application demonstrates successful cross-pollutant knowledge transfer using artificial neural networks, where models pre-trained on PM2.5 data are effectively fine-tuned to predict other pollutants, enabling efficient modeling in data-scarce scenarios. The second application extends temporal transfer learning from hourly to multi-horizon forecasting, allowing models to adapt across different prediction timescales with minimal computational overhead. The third application introduces our novel LagEnsembleForecasting framework, which combines four specialized deep learning models trained on different temporal lag sequences (1h, 6h, 12h, 24h) to provide enhanced prediction accuracy and natural uncertainty quantification through ensemble variance, with successful transferability to other pollutants demonstrating the framework's versatility.Third, we advance particulate matter risk assessment through machine learning-driven toxicity threshold prediction, addressing the limitations of costly and time-intensive laboratory approaches. By leveraging a comprehensive dataset compiled from over 40 peer-reviewed studies, our framework employs five complementary algorithms with SHAP-based explainability to distinguish toxic from non-toxic PM concentrations. This approach provides more nuanced toxicity assessments than conventional methods while identifying key compositional factors that drive toxicity outcomes.The integration of these contributions establishes a comprehensive framework for intelligent environmental monitoring that enhances prediction capabilities, provides interpretable decision support, and enables more effective health risk assessment. Validation across multiple real-world datasets confirms the practical applicability and scalability of our approaches. Importantly, while this research focuses primarily on air quality, the methodological foundations and transfer learning principles established here create a pathway toward integrated water-air-soil monitoring strategies, where the same intelligent frameworks could be adapted and extended to create unified environmental management systems that holistically address the interconnected nature of environmental challenges across multiple domains.