This study analysed the effect of substituting different percentages of natural aggregate with recycled aggregate from concrete crushing, using a coarse fraction as well as a fine fraction. Natural and recycled materials were classified in order to analyse the mechanical performance and impermeability of these eco-concretes in the fresh state as well as in the hardened state. A statistical analysis also determined whether the performance loss was significant from a statistical point of view, finding strength decreases of less than 13% in compressive strength and losses of less than 20% in flexural strength. An increasing trend was found in permeability as the percentage of recycled aggregate in the mix increased.
Construction and demolition waste (CDW) management remains a significant challenge due to the large volume and heterogeneity of waste streams in real-world recycling facilities. This work addresses automated CDW classification under mixed-material conditions using deep learning. From an artificial intelligence perspective, a global multilabel image classification approach based on transfer learning with a 50-Layer Residual Network (ResNet50) is proposed as a computationally efficient alternative to detection and segmentation methods in scenarios with strong material overlap and limited annotation. Gradient-weighted Class Activation Mapping (Grad-CAM) is further integrated to provide visual interpretability and support analysis of model behaviour. A representative dataset was collected at a recycling plant for model training. From an engineering application perspective, the system is designed for deployment in material recovery facilities, implemented as a web-based prototype capable of processing both static images and real-time video. Experimental results demonstrate reliable performance under realistic conditions, achieving a macro‑F1 score of 0.989 on mixed‑material CDW images, and the system is currently in the final stage of implementation for deployment in an operational CDW treatment facility. These findings support the use of interpretable deep learning models for scalable and practical CDW management.