An advanced AI-driven approach to burned surface quantification in fire performance testing of wood materials

Reproducible quantification of fire-induced surface degradation remains challenging in laboratory testing of wood-based materials. This study developed and validated an AI-assisted workflow for time-resolved quantification of burned surfaces combining computer vision, manual mask refinement, and deep-learning segmentation. Sequential red, green, and blue light values (RGB) images from controlled fire tests of lignocellulosic insulation fiberboards were geometrically corrected and processed to generate burned-area masks. These masks were manually refined to remove systematic errors and used to train a U-Net model with a ResNet34 encoder. Four approaches were compared: raw computer vision (CV), manually corrected masks (CV-manual), AI trained on raw CV data (AI-CV), and AI trained on refined data (AI-CV-manual). Performance was evaluated using Intersection over Union (IoU), Dice score, and a temporal pixel-consistency criterion. The U-Net models achieved IoU values up to 0.9961 and Dice scores up to 0.9980. Manual refinement reduced the average temporal error from approximately 1.1% to 0.6%, while AI-CV-manual achieved the lowest error of approximately 0.5%. The results demonstrate the importance of reference annotation quality for reliable deep-learning segmentation. The proposed workflow provides a reproducible and scalable approach for spatially resolved monitoring of burned-surface development in wood-based materials.