Evaluating Post-Hoc Explainable AI Methods for a Multitask-CNN in Sensor-Based Quality Monitoring of Comminuted Construction and Demolition Waste

Abstract

There is a lack of comparative XAI evaluations for sensor-based quality monitoring (SBQM) appli-cations in the context of construction and demolition waste (CDW) management. Moreover, such studies are generally oriented towards classification architectures and tasks. To address these gaps, we compared the faithfulness, stability and localization of three common post-hoc XAI evaluation methods, namely Grad-CAM, LIME and SHAP. These post-hoc methods were implemented into a multitask-CNN, i.e. CDW-MT-CNN, containing an identical classification and regression head architecture. No single method excels on all aspects: only Grad-CAM is fast and stable enough for near-real-time monitoring (10 ms/instance) but suffers from its large receptive field limitation, impacting its faithfulness and localization. LIME and SHAP are faithful and localize well but have limited stability and their computational overhead (180–600 ms/instance) renders them unsuitable for the near-real-time application. Further research in the directions of correlation studies between explanation and predictive quality, a detailed sensitivity study of the hyperparameters for LIME and SHAP, explainability evaluations of advanced gradient-based methods that mitigate the large receptive field limitation, and a detailed per-image runtime and energy consumption study on representative recycling plant hardware would further contribute to responsible deployment of XAI in SBQM for CDW management.