Contextual fusion of classifiers under variable atmospheric conditions for coastal surveillance

Monitoring maritime areas is a challenging task which generally benefits from sensing systems such as radars or cameras, with different electromagnetic or optical capabilities. Infrared (IR) cameras in particular allow night and day surveillance, but their image is affected by maritime atmospheric conditions, e.g., the aerosol concentration. Further classification process thus inherits this image degradation resulting in possibly poor target classification results. To overcome this issue, we propose in this paper a contextual classifiers fusion system where two neural networks are trained into two environmental contexts and further combined with Bayesian reasoning. Indi-vidual classifier's reliability is considered, enabling to balance between the two classifiers depending on the uncertain context of use. Additionally, we apply an imprecise decision rule for a greater flexibility allowing a compromise between two criteria of accuracy and specificity. Results are obtained on simulated classifiers' outputs as well as on a synthetic dataset of IR images degraded with atmospheric context. It is shown that the proposed approach allows the possibility to increase accuracy while foregoing specificity and adds explainability.