Lightweight deep learning for sustainable adobe construction: a block-level baseline aligned with circular economy principles and sustainable development goals
Journal
Frontiers in Built Environment
ISSN
22973362
Date Issued
2026-07-08
Type
Article
Abstract
Introduction – The construction sector accounts for nearly 37% of global energy-related CO2 emissions, motivating renewed interest in low-carbon vernacular materials such as stabilized adobe. The integration of recycled inputs into adobe formulations advances circular-economy principles in the built environment, yet large-scale adoption is constrained by the absence of objective, scalable quality-control methods. Methods – We present a methodologically rigorous baseline for AI-driven visual quality control in stabilized adobe construction. A dataset of 330 images was constructed from 55 physical adobe blocks fabricated under controlled laboratory conditions with three circular-economy stabilization mixtures: recycled polyethylene terephthalate (1.5%), sugarcane bagasse fiber (8%), and Portland cement (10%). Each block was photographed from six standardized angles. To eliminate data leakage, we implemented strict block-level data splitting. A MobileNetV2 transfer learning pipeline with frozen ImageNet weights, aggressive data augmentation, L2 regularization, and balanced class weights was developed. Performance was characterized through accuracy with bootstrap 95% confidence intervals, per-class metrics, ROC AUC and Average Precision, five-fold block-level cross-validation, ablation study, and Grad-CAM interpretability. Results – The model achieved 80.95% test accuracy (95% CI: 69.0%–92.9%) on a held-out set of seven physical blocks (42 images). AUC = 0.858 and Average Precision = 0.944. Per-class F1-scores were 0.692 (sound) and 0.862 (defective). Block-level five-fold cross-validation confirmed stability at 70.0% (95% CI: 60.7%–79.2%). Ablation showed: simple CNN 71.43%, MobileNetV2 without augmentation 71.43%, complete pipeline 80.95% (+9.5 pp). Grad-CAM confirmed attention to physically meaningful surface features. The 9.24 MB model footprint enables offline deployment on commodity smartphones. Conclusion – This work establishes a methodologically rigorous baseline for AI-driven quality control in stabilized adobe construction. By demonstrating meaningful performance with modest datasets and edge-deployable architectures, this study supports broader adoption of circular-economy adobe construction and contributes to Sustainable Development Goals 9, 11, 12, and 13. Future work should expand the dataset, evaluate hierarchical classification, and assess field deployment. Copyright © 2026 Pinto Almeida and Jordán Buenaño.
Conclusion
This work establishes a methodologically rigorous baseline for AI-driven quality control in stabilized adobe construction. By demonstrating meaningful performance with modest datasets and edge-deployable architectures, this study supports broader adoption of circular-economy adobe construction and contributes to Sustainable Development Goals 9, 11, 12, and 13. Future work should expand the dataset, evaluate hierarchical classification, and assess field deployment.
Conclusion
This work establishes a methodologically rigorous baseline for AI-driven quality control in stabilized adobe construction. By demonstrating meaningful performance with modest datasets and edge-deployable architectures, this study supports broader adoption of circular-economy adobe construction and contributes to Sustainable Development Goals 9, 11, 12, and 13. Future work should expand the dataset, evaluate hierarchical classification, and assess field deployment.