New joint estimation method for emissivity and temperature distribution based on a Kriged Marginalized Particle Filter: Application to simulated infrared thermal image sequences

• DOI: 10.1016/j.srs.2025.100209
• Publication Date: 2025-02-22T12:10:56Z
• Authors: Thibaud Toullier; Jean Dumoulin; Laurent Mevel
• Keywords: Thermal infrared; Particle (ecology)

This paper addresses the challenge of simultaneously estimating temperature and emissivity for infrared thermography in natural environment, aiming for near real-time performance. Existing methods, mainly in satellite observation field, rely on restrictive physical assumptions unsuitable for ground-based application context (Structures and Infrastructures monitoring). Other generic methods are nonetheless computationally intensive, making them impractical for real-time use. Our objective is to provide a method with effective real-time calculation performance while still giving results comparable to those reference methods under the same hypotheses, finally achieving both good accuracy and performance. The proposed method is based on a dynamical state-space modeling for the temperature, where the state vector is assumed to be split into a dynamic component for the temperature and a stationary component representing the emissivity. Then the dynamical component is estimated by a Kalman filter approach, whereas the parameterized model and the emissivity component are estimated through a particle filtering framework resulting in a bank of Kalman filters, also called marginalized particle filter. A spatial assumption of homogeneity for the temperature yields to the addition of a Kriging step to the Marginalized Particle Filter to overcome the ill-posed nature of the problem and to compute the necessary physical estimates in a reasonable amount of time while providing fair results compared to reference methods from the literature.A comparison with two state-of-the-art methods, MCMC and CMA-ES, is presented. The results indicate that the proposed method estimates the true value within a maximum deviation of 3K, similar to CMA-ES, while MCMC achieves a more accurate estimate with a maximum deviation of 0.5K. However, the computational efficiency of the proposed method is significantly improved, reducing the processing time by seven orders of magnitude compared to MCMC and three orders of magnitude compared to CMA-ES. This remarkable efficiency highlights the method’s feasibility for real-time monitoring of temperature and emissivity.