Early detection of defects (e.g., cracks, delamination in composites, defects in adhesive bonds) is critical to prevent potential accidents in industrial environments. In this context, nondestructive evaluation or testing (NDT/E) is essential to avoid such disastrous failure. Among nondestructive testing techniques, sonothermography (also called vibrothermography) comprises coupling powerful ultrasounds and thermal imaging to induce material defects and evaluate the extent of the damage. In the presence of a defect, the mechanical energy induced by the ultrasonic waves is converted into thermal energy, resulting in the buried defect being converted into a thermal source. By studying the resulting thermal and mechanical fields, inverse methods have been developed to reconstruct these heat sources, enabling the evaluation of physical properties for materials or defects. Here, we focus on the reconstruction of volumic sources in material regardless of their origin from thermal imaging registered at the surface of the material using a Bayesian approach denoted "GB". This problem is notably ill-posed and requires the use of specific tools to ensure this reconstruction from noisy experimental data. We consider prior modelling in the form of a mixture of distributions, which promote spatial sparsity and thereby regularize the inference problem. This method enables the reconstruction of sources with high fidelity, including the estimation of this intensity, thus overcoming the limitations mentioned in previous works. Its performance is discussed in relation to noise, and it is compared to other methods, showing favourable potential for NDT/E applications.Read less <