Méthodes statistiques pour le déploiement de capteurs innovants en environnements non-contrôlés et en présence d'incertitudes

Air and water pollution are major public health concerns that highlight the need for reliable and affordable environmental monitoring tools. In this context, this thesis aims to develop robust calibration methods for sensors that should be well adapted to real-world, uncontrolled environments that are inherently variable and affected by numerous sources of uncertainty. Although innovative sensors offer promising advantages, such as low cost and high sensitivity, their practical deployment still faces several obstacles. In particular, the relationships between sensor signals and environmental variables are often complex, nonlinear, and unknown a priori. Furthermore, a moderate selectivity, strong correlations between covariates, measurement noise, and the presence of unobserved variables complicate both model calibration and the identification of truly influential variables. These difficulties call for careful modelling choices to balance bias and variance in estimation.To tackle these issues, the thesis first proposes a Bayesian calibration framework that explicitly accounts for multiple sources of uncertainty, including uncertainty in the input variables, an aspect often overlooked in standard approaches. The method combines a non-parametric model, to flexibly capture complex dependencies, with a parametric model, in order to exploit the structure of the sensor responses. This hybrid approach enables robust calibration even in the presence of noisy, correlated, or incomplete data.In addition, the thesis introduces a method for identifying variables that have non-negligible influence on the sensor outputs. This approach aims at selecting a minimal yet sufficient subset of these variables to ensure reliable pollutant prediction. The approach shows robustness to noise, correlation and unobserved variables, while providing interpretable and consistent selection criteria.These methods have been applied to both simulated and real-world sensor deployment datasets. These applications demonstrate the effectiveness and practical applicability of the proposed methods in realistic settings. Overall, this work contributes to the development of calibration tools that are both practical and theoretically grounded, paving the way for more robust and reliable environmental monitoring using innovative low-cost sensors in real-world applications