Explainable AI for shooting distance estimation from gunshot residue (GSR) via flash thermography

Shooting distance estimation from a gunshot residue pattern (GSR) is traditionally performed using chemical and optical methods. Infrared thermography offers a non-destructive alternative for visualising GSR patterns on textile substrates. This work proposes a data-driven approach to estimating shooting distance by analysing thermographic images with a convolutional neural network (CNN). Using kurtosis images were selected as a compact, high-contrast representation of thermographic sequences. A large experimental dataset covering a wide range of shooting distances, firearms, and ammunition types was analyzed. Several CNN architectures were evaluated using transfer learning and fine-tuning. A multimodal architecture combining image features with categorical ballistic information was also investigated. A VGG-based model achieved the most reliable performance, while the multimodal model showed overfitting, suggesting potential for larger, more diverse datasets. Model explainability was addressed through error analysis and Grad-CAM visualisations, confirming that the networks focus on physically meaningful regions. The proposed approach demonstrates the potential of combining infrared thermography with deep learning for reliable, robust forensic estimation of shooting distance.