Integrating automation into construction waste quantification: A review of digital innovations for sustainable resource management

Abstract

The construction sector is a major generator of solid waste, with construction and demolition waste (CDW) posing significant challenges to sustainable resource management. Accurate quantification is vital for achieving circular economy and waste reduction goals; however, traditional estimation methods remain manual, fragmented and inconsistent. This state-of-the-art review synthesizes developments in automation and digital technologies that are transforming CDW quantification and management. A total of 125 peer-reviewed articles published between 1993 and 2025 were systematically analysed to identify trends, methods and emerging tools. Advances across artificial intelligence, computer vision, building information modelling and the Internet of Things are categorized, focusing on automated waste recognition, volumetric estimation and real-time monitoring. Persistent challenges, including limited dataset diversity, model generalization, interoperability and implementation cost, are critically examined. By linking technological innovation with practical waste management, this review highlights how automation can enhance efficiency, traceability and sustainability in the construction industry’s transition towards circular practices.

Introduction

The construction sector plays a vital role in economic growth and urban development but remains one of the largest generators of solid waste worldwide. Construction and demolition waste (CDW) often accounts for nearly one-third of total solid waste in many regions, posing major challenges to sustainable waste management and resource recovery (Chen et al., 2018; Devaki and Shanmugapriya, 2022; Ram and Kalidindi, 2017). Achieving circular economy (CE) and waste minimization targets depends heavily on the accuracy with which CDW is measured, tracked and reported. Reliable quantification is critical to identifying recoverable materials, reducing on-site waste generation and improving recycling efficiency. However, conventional estimation methods continue to fall short of providing the precision, scalability and transparency required for modern waste management systems (Cudecka-Purina et al., 2024; Dodampegama et al., 2024; Quiñones et al., 2022). Global variations in CDW generation and recovery highlight the need for improved quantification practices (López Ruiz et al., 2020). As shown in Figure 1, China generates approximately 2360 million tonnes of CDW but recovers only 5%, while the United States produces around 534 million tonnes with a 70% recovery rate. The EU-28 generates about 924 million tonnes, with a 46% recovery rate, and the United Kingdom achieves one of the highest rates at 90% from 136 million tonnes (Gatto, 2023; Joseph et al., 2023; Mercader-Moyano and Ramírez-De-Arellano-Agudo, 2013). High-performing countries benefit from standardized measurement protocols and data-driven reporting systems, whereas others struggle due to the absence of robust waste-quantification frameworks (Wang et al., 2019).

Figure 1.

Country-wise CDW: (a) annual production and (b) recovery rates (reproduced from López Ruiz et al. (2020) with permission, License number: 6142910695513).

CDW: construction and demolition waste.

These disparities demonstrate that even with advanced recycling infrastructure, waste recovery can stagnate when reliable measurement systems are lacking. Accurate, material-level, real-time quantification is therefore a prerequisite for circular construction and effective waste governance. The ability to measure waste precisely depends not only on policy and infrastructure but also on the tools and technologies used. Traditional approaches, such as per capita multipliers, generation-rate models and project-based waste indices, remain low cost and easy to apply but are typically manual, time-consuming and limited in accuracy (Masudi et al., 2011; Wu et al., 2014). They often fail to capture mixed waste streams, site variability and dynamic changes during construction, leading to inconsistent data and unreliable recovery estimates. Recent advances in digitalization and automation are redefining waste measurement and management. Building information modelling (BIM) enables early-stage waste forecasting and material optimization (Quiñones et al., 2022), while sensing technologies such as RGB/RGB-D imaging, hyperspectral analysis, unmanned aerial vehicle (UAV) photogrammetry and Internet of Things (IoT) systems allow near-real-time data collection from active sites. When integrated with artificial intelligence (AI) and computer vision (CV), these technologies can automatically identify materials, estimate waste volumes and monitor recycling performance. Despite encouraging progress, several challenges remain. Many AI-driven studies are still limited to controlled environments, rely on small datasets and focus primarily on waste classification rather than on volumetric estimation (Dodampegama et al., 2024). Moreover, standardized evaluation metrics such as precision, recall and mean intersection-over-union (IoU) are inconsistently applied, hindering the comparability of models. Integration of quantification outputs into operational workflows, such as BIM-based project management, remains limited (Dodampegama et al., 2024; Quiñones et al., 2022).

This review presents a state-of-the-art synthesis of automation and digital technologies applied to construction waste quantification. It critically examines 125 peer-reviewed studies to identify emerging trends, technological progress and implementation barriers. The paper bridges academic innovation and practical waste management by evaluating how automation improves accuracy, traceability and circular resource use. Emphasizing the roles of AI, BIM and IoT, it highlights their potential to create standardized, scalable systems that support predictive analytics, policy development and sustainable material recovery. Overall, the review advances understanding of data-driven, automation-enabled practices that can enhance efficiency and circularity in construction waste management.

Scientometric analysis

A scientometric analysis using the Biblioshiny software on the keyword “construction waste sorting” reveals steady and accelerating growth in research over the past three decades, reflecting the construction sector’s increasing commitment to sustainability and resource efficiency (Aria and Cuccurullo, 2017). Figure 2 presents the research workflow and data extraction process used in this scientometric assessment. The study employed keyword searches in Scopus to identify 125 publications from 1993 to 2025.

Figure 2.

Flowchart for research methodology.

To ensure thorough and repeatable coverage beyond a single keyword query, the literature search was carried out using a structured combination of keywords, “construction waste sorting”, “demolition waste”, “waste management”, “automated sorting”, “artificial intelligence” and “automations,” applied with Boolean operators. After applying inclusion criteria (i.e. peer-reviewed English-language studies addressing automated or technology-assisted construction waste-sorting methods) and excluding editorials, duplicates and those lacking methodological detail, the final dataset comprised 125 articles for a detailed review. The search approach follows the PRISMA-type flow (Moher et al., 2009), as illustrated in Figure 3.

Figure 3.

PRISMA-flowchart for selection criteria in database.

As shown in Figure 4, research output remained low in the early years but has expanded rapidly over the past decade, driven by stricter waste management policies and the broader adoption of CE principles. Frequently co-occurring keywords, such as waste management, recycling, automation, artificial intelligence, sorting and computer vision, illustrate the field’s shift towards technologically advanced and data-driven solutions, as shown in Figure 5.

Figure 4.

Number of articles annually published; keyword: “Construction waste sorting”.

Figure 5.

Co-occurrence network of the keyword “Construction waste sorting”.

A thematic analysis of keyword clusters, illustrated in Figure 6, reveals a clear shift in research trends from early studies focusing on manual and semi-automated sorting to recent emphasis on CV, robotics and AI-based automation for optimizing waste sorting. Country collaboration mapping identifies China, the United States, the United Kingdom, Australia and Germany as leading contributors, with particularly strong partnerships between European countries, Australia and China, as shown in Figure 7. These collaborations have advanced waste-to-value approaches and automated sorting methodologies.

Figure 6.

Thematic evolution of the keyword clusters.

Figure 7.

Country collaboration map.

Table 1 summarizes major regional projects and studies, highlighting their key drivers, outcomes and limitations for waste sorting and automation. The automated CDW systems differ significantly in their drivers and levels of maturity when regional research and implementation trends are compared. Research expansion in the EU is heavily influenced by policy, supported by legally binding CDW recycling objectives, circular-economy plans and ongoing public financing through extensive cooperative initiatives like Horizon Europe. Therefore, system-level integration, interoperability and pilot-scale validation across real-world building and recycling contexts are prioritized in EU efforts, such as RECONMATIC, MOBICCON-PRO, RECONSTRUCT and RISEN (Cordis, 2022; Horizon Europe, 2023; Kaewunruen, 2016; Mobiccon-Pro, 2023). However, integration with legacy infrastructure and the shift from pilot demonstrations to full industrial deployment are common problems reported by these programs. Recent research shows that CDW management automation delivers a substantial return on investment (ROI) for Small and Medium-sized Enterprises (SMEs) by lowering labour costs, increasing accuracy and enabling scalable waste management solutions, although the hefty CAPEX and deployment costs might be a barrier (Farshadfar et al., 2025; Oliveira Neto et al., 2017). On the other hand, methodological frameworks, conceptual automation pipelines and partial digitalization through BIM and sensing are the main topics of research in the United States, which is primarily driven by professional associations and industry-academia partnerships (Alwadhenani et al., 2022; Gibson, 2020; Kamma and Jha, 2022). Despite being technologically competent, many US-based research projects remain conceptual or semi-automated, and fewer large-scale robotic sorting implementations are documented. This is indicative of a less organized regulatory drive than in the EU. In the meantime, Asian research, especially from China and the East Asian region, shows quick progress in edge/cloud computing architectures, multi-robot coordination and CV accuracy, frequently attaining high performance under regulated circumstances (Dong et al., 2022; Hondo et al., 2020; Liu et al., 2022). Although national smart construction and AI programs promote these breakthroughs, extremely diverse waste streams and insufficient long-term empirical evidence sometimes hinder practical deployment. Conclusively, the comparison shows that regulatory compliance as well as system integration requirements shape EU research, standards and methodological development shape US research, and rapid engineering optimization shapes Asian studies. These factors collectively account for the noticed regional variations in research findings and implementation goals.

Table 1.

Regional perspectives for waste sorting and automations.

Initiative (region)	Key drivers	Outcomes	Challenges	References
RECONMATIC (EU Horizon Europe)	• EU circular economy strategy • Horizon Europe funding for automated CDW valorization • Regulatory pressure for higher recycling rates	• AI-based CDW classification and tracking • Multi-sensor data fusion (vision, spectroscopy) • Pilot robotic sorting cells	• Industrial-scale validation ongoing • Limited long-term operational data • Integration with existing facilities under assessment	Cordis (2022); Koroxenidis et al. (2023)
MOBICCON-PRO (EU Horizon Europe)	• EU policy on selective demolition • Territorial circularity objectives • Public research funding	• Mobile selective separation units • Digital tools for waste flow management • Regional pilot demonstrations	• Broader focus than pure automation • Automation performance metrics limited • Interoperability issues remain implicit	Cousture et al. (2025); Mobiccon-Pro (2023)
RECONSTRUCT (EU Horizon)	• EU Green Deal • Resource efficiency mandates • Research funding for reuse of CDW	• Digital tools for material reuse planning • Automated material identification concepts • Demonstration on real construction assets	• Focus primarily on reuse planning • Limited robotic sorting validation • Data harmonization challenges	Calcagni et al. (2025); Horizon Europe (2023); Mobili et al. (2024)
RISEN (EU Horizon Europe)	• Resilient infrastructure policy • Sustainability and circularity funding	• Integrated monitoring and digitalization • Lifecycle-based CDW assessment methods	• Automation is indirect (decision-support) • Limited focus on on-site sorting hardware	Kaewunruen (2016); Kaewunruen et al. (2021)
SmartRecycling-Up (Germany)	• National innovation funding • Industrial AI and robotics programs	• AI-based robotic sorting of bulky waste • Sensor integration with heavy equipment • Pilot-scale operational testing	• High capital and integration cost • Complex system setup • Limited scalability evidence	Lange et al. (2023); Vögele et al. (2025)
BAM Berlin (Germany)	• Federal research funding • Regulatory compliance research	• Experimental studies on automated material identification • Validation of sensor-based sorting concepts	• Focus on experimental validation • Limited full-scale industrial deployment	Klewe et al. (2022); Wilsch and Völker (2024)
Christian Doppler Lab (Vienna)	• Industry-academia collaboration • National research funding	• Advanced CDW characterization methods • Automation-oriented waste processing research	• Primarily laboratory and pilot studies • Scaling challenges to heterogeneous CDW	Kleemann et al. (2017); Lederer et al. (2021)
CDW policies (USA)	• Sustainability standards • Industry-driven research	• Frameworks for automated CDW quantification • Integration concepts with BIM and sensing	• Mostly conceptual or semi-automated systems • Limited full robotic sorting cases	Alwadhenani et al. (2022); Gibson (2020); Halevi et al. (2023); Yost and Halstead (1996)
Vision-based Robotic Sorting (Australia)	• AI and robotics research funding • Automation demand in waste sector	• High classification and grasping accuracy in lab settings • Deep learning for classifying C&D wood waste • Open annotated datasets	• Strong dependence on labelled data • Performance drops in real heterogeneous waste	Gautam et al. (2026); Ranjbar et al. (2026); Sirimewan et al. (2025)
Industrial Pilot Plants (EU and Asia)	• Regulatory landfill diversion targets • Industrial innovation incentives	• Demonstrated efficiency gains in pilot facilities • Reduced manual sorting effort	• High CAPEX and OPEX • Skilled workforce requirement • ROI uncertainty for SMEs	Hondo et al. (2020); Huang et al. (2018); Liu et al. (2022)
Market-driven Automation Systems	• Commercial waste management demand • Compliance with recycling quotas	• Increased interest in robotic CDW sorting • Gradual pilot adoption	• Hardware cost barriers • Integration with legacy systems • Maintenance complexity	Farhadi (2024); Liu et al. (2025); European Commission (2016)

CDW: construction and demolition waste; AI: artificial intelligence; BIM: building information modelling; ROI: return on investment; SMEs = Small and Medium-sized Enterprises.

The growing integration of digital technologies reflects a move towards data-driven strategies, enabling real-time monitoring and optimization of sorting processes. Thematic clusters also emphasize the field’s multidisciplinary nature, combining AI, environmental science and civil engineering. In line with Industry 4.0, this trajectory supports global efforts to promote sustainable construction and intelligent waste systems. AI plays a pivotal role in developing waste management models that integrate social, economic and environmental factors, particularly in the transition towards a ‘zero-waste’ CE (Farghali and Osman, 2024). The importance of accurate waste quantification lies in three main areas:

1. Data-driven decision-making: Reliable quantification supports targeted waste reduction, enhanced recycling and optimized resource utilization (Jiang et al., 2020; Kumar et al., 2021). AI-assisted analytics have revolutionized recycling by improving collection, processing and resource allocation workflows (Midigudla et al., 2025). For example, Jiang et al. (2020) applied data mining and IoT techniques to model household waste disposal behaviours and inform policy.

2. CE implementation: Measurement according to CE standards identifies recyclable materials, reuse opportunities and value-added applications (Moraga et al., 2019; Yang et al., 2023). Du et al. (2024) further explored how next-generation digital technologies strengthen the CE framework.

3. Environmental impact reduction: Accurate measurement helps organizations identify ways to reduce pollution, resource depletion and greenhouse gas emissions, supporting more sustainable waste practices (Chang et al., 2011).

Various recent studies report that AI-based sorting systems can significantly reduce contamination and improve recovery efficiency, although they still face challenges in recognizing complex, mixed or contaminated inputs in real time (Cha et al., 2022; Lakhouit, 2025; Lin et al., 2024; Midigudla et al., 2025; Radhakrishnan Nair et al., 2024). The integration of deep learning (DL), CV and IoT has transformed resource recovery workflows, enabling higher accuracy, efficiency and scalability over traditional techniques, and is thoroughly reviewed in the forthcoming sections.

Traditional techniques for waste quantification

Historically, the construction sector has relied on a range of manual and conventional techniques to quantify CDW (Colorado et al., 2022; Lau Hiu Hoong et al., 2020; Li and Yang, 2023; Prochazka et al., 2024; Rašković et al., 2020; Xiao et al., 2020). These methods, illustrated in Figure 8, have evolved alongside changing construction practices and regulatory demands, but continue to face notable limitations in accuracy, scalability and timeliness compared with emerging AI-based approaches. For example, one study compared manual image recognition, DL-based recognition and manual sorting, demonstrating significant efficiency and accuracy gains through automation (Jiang et al., 2025).

Figure 8.

Conventional methods for construction waste identification and sorting.

Common manual techniques include sorting and weighing, site audits and material flow analysis. Sorting and weighing involve manually categorizing and measuring waste to obtain detailed material-level data; however, this process is slow, labour-intensive and prone to human error. Site audits estimate waste generation using disposal records and checklists during regular site visits, providing data to develop a building waste index. Material flow analysis monitors material movement throughout construction to identify waste generation points and gain a better understanding of production dynamics (Hassan et al., 2018; Mahayuddin and Zaharuddin, 2013; Masudi et al., 2011).

Traditional CDW estimation approaches include the per capita multiplier, waste generation rate model and Construction Waste Index (Masudi et al., 2011; Umar et al., 2017; Wu et al., 2014; Yost and Halstead, 1996). The per capita multiplier estimates CDW from population and average generation rates, but is often unreliable due to regional variation. The waste generation rate model multiplies construction quantities by material-specific generation rates, requiring comprehensive and accurate datasets. The Construction Waste Index relies on historical project data to forecast waste generation and raise awareness of management, although its precision depends on the relevance of past conditions.

Manual and conventional quantification methods share several persistent drawbacks (Hassan et al., 2018; Masudi et al., 2011). They are time- and labour-intensive, relying heavily on on-site visits and manual data recording, which delays decision-making. The lack of standardized procedures makes cross-project comparisons difficult, as waste generation varies with project type, location and material composition. Estimates often rely on subjective or outdated data, which limits their real-time applicability. Furthermore, these methods typically address only specific materials or project stages, overlooking total waste generation. These limitations highlight the growing need for automation and AI-driven solutions. AI and robotics can enhance waste processing, on-site sorting and recycling efficiency by learning from project-specific data to improve precision, predict waste generation and optimize workflows (Farghali and Osman, 2024; Zia et al., 2025). Advanced computational techniques, including machine learning (ML), DL, artificial neural networks (ANN), genetic algorithms and fuzzy logic, are particularly effective in modelling complex and uncertain relationships beyond the capacity of traditional methods. On-site sorting practices in countries such as Japan, where strict regulations have driven recycling rates above 90%, demonstrate the benefits of automation, including reduced storage and transportation costs and lower environmental impacts. Nonetheless, space constraints, administrative complexity and high labour costs continue to limit widespread adoption, underscoring the need for scalable and cost-effective automation technologies (Chen et al., 2022b; Dodampegama et al., 2024).

Digital imaging in construction waste quantification

Digital imaging has emerged as a vital tool for quantifying construction waste, offering greater precision than human visual assessment. While human vision relies on size, shape, colour, texture and occasionally weight to identify materials, it is limited in its scope. Imaging technologies overcome these constraints by enabling multi-attribute detection and monitoring with high accuracy. They now play a crucial role in managing and quantifying construction waste, reducing manual effort and promoting sustainability through improved recycling and waste reduction (Chen et al., 2022b; Dodampegama et al., 2024). Before the advent of DL, waste classification depended primarily on traditional ML approaches such as the Viola–Jones object detection framework, scale-invariant feature transforms and histograms of oriented gradients (Arlazarov et al., 2021; Viola and Jones, 2001). Algorithms, including random forests (RF), logistic regression and colour histograms, grouped image features into clusters of similar attributes. ML enabled pattern extraction from structured and unstructured datasets, but DL, powered by high-performance Graphics Processing Units (GPUs), has since transformed the field (Abdu and Mohd Noor, 2022). Convolutional neural networks (CNNs) and related DL architectures now allow real-time identification and classification of waste materials, supporting accurate separation and quantification (Olawade et al., 2024). Recent studies illustrate these advances. For instance, one study used 3D imaging and laser scanning to track waste during deconstruction, addressing challenges such as limited visibility and uncertainty in appearance (Wei et al., 2019). Another study demonstrated that a BIM-based tool (WE-BIM Add-in) in Revit can estimate waste volumes during the design phase, enabling early waste-reduction strategies (Quiñones et al., 2022). A third study highlighted the importance of integrating digital technologies, including blockchain, BIM, AI, robotics and CV, to enhance waste control and sustainability (Iyiola et al., 2024).

A range of imaging techniques is now used in CDW classification. RGB cameras are commonly used to capture video data of metals, polymers and wood, providing essential colour information for material separation (Dodampegama et al., 2024). Enhanced systems, such as RGB-D, RGB-D point clouds, hyperspectral and thermal imaging, capture depth, spectral signatures and thermal properties. For volumetric estimation and 3D scanning, both RGB and monochrome imaging offer material-specific identification (Chen et al., 2022b; Muri and Hjelme, 2022; Rašković et al., 2020). Drone-based photogrammetry further expands these capabilities, allowing aerial mapping, waste scanning and automated volume estimation with DL algorithms (Anadkat et al., 2019; Jiang et al., 2022). Table 2 summarizes the main sensor types, data formats, extracted features and their advantages and limitations. Different imaging systems offer distinct advantages and trade-offs. RGB sensors are cost-effective and easy to deploy, but struggle with shadow effects and overlapping materials. Hyperspectral imaging provides detailed chemical composition data but entails higher equipment and processing costs. Hybrid systems that combine multiple sensing modalities can achieve greater accuracy but face challenges in data alignment and computational demand (Dodampegama et al., 2024). Integrating these imaging modalities with AI enables waste quantification to shift from static, periodic assessments to continuous, near-real-time monitoring. This transition supports the incorporation of accurate, verifiable waste data into BIM-based planning and CE metrics, directly addressing the limitations of traditional manual methods.

Table 3.

Construction and demolition waste sorting and segmentation analysis (Langley et al., 2025).

CDW collection setting	Materials examined	Camera/sensor type	Primary task/method(s)	Image resolution	Dataset characteristics	References
Construction site	Common CDW such as concrete, brick, stone, ceramic tile, glass, metal scrap, gypsum board, wood, plastic and paper.	RGB	Object classification utilizing ResNet-based models.	Not disclosed (images scaled to 224 × 224 pixels).	2836 original images, augmented to 36,711 for training, 4080 for validation and 10,203 for testing.	Lin et al. (2023)
	Construction materials and waste, including concrete, brick, stone, ceramic tile, glass, metal scrap, gypsum board, wood, plastic and paper.	RGB	Object classification using VGG-based models.	Not disclosed (images scaled to 224 × 224 pixels).	2836 original images, augmented to 36,711 for training, 4080 for validation and 10,203 for testing.	Lin et al. (2022, 2023)
	Concrete brick, board, mixed waste and wood.	RGB	Object classification, detection, and segmentation using YOLACT.	512 × 512 pixels.	599 images, sourced from construction sites and web crawling.	Sunwoo et al. (2022)
	Brick, wood, stone and plastic waste.	RGB	Object detection using Faster-RCNN and YOLO.	Not disclosed	3046 images; 80% for training, 10% for testing and 10% for validation.	Zhou et al. (2023)
	Cotton gloves, wood blocks, small ferrous metals, plastic pipes, bamboo, corrugated paper and rebar.	RGB-Depth, 3D LiDAR	Object detection and segmentation via Mask R-CNN.	640 × 480 pixels.	756 images; 454 for training, 151 for validation and 151 for testing.	Chen et al. (2022b)
	Concrete brick, lumber, board and mixed waste.	RGB	Object detection and segmentation with YOLACT.	512 × 512 pixels.	788 images (500 from construction site, 288 from web crawling).	Na et al. (2022)
	Construction debris including rebar, bricks, PVC pipes, plastic wires and general debris.	RGB	Object segmentation applying Swin Transformer, Twins Transformer and K-Net.	Average resolution of 620 × 770 pixels.	1919 images (including synthetic data); 1696 for training, 200 for testing and 23 for validation.	Wang et al. (2023)
	Rebar, bricks, PVC pipes, wires and cementitious debris.	RGB	Object segmentation utilizing Detectron2, YOLACT and MMDetection.	Not disclosed.	858 images.	Ko et al. (2022)
Conveyor belt	Concrete grains, natural stones, ceramics, bituminous grains, glass and other materials.	RGB	Multi-task approach: object classification, segmentation and mass regression using RACNET.	8192 × 4096 pixels (from Database 2).	Images derived from 2 datasets; 8000 for training, 800 for testing and 800 for validation.	Lux et al. (2023)
	Bricks, wood, plastics, concretes and foams.	RGB	Object classification utilizing VGG-16, ResNet-50 and Transformer models.	Not disclosed.	125 pictures per class; 100 samples for training and 25 for testing per material type.	Song et al. (2022)
	Wood, rubber, brick and concrete.	Near-Infrared (NIR)hyperspectral	Object classification using a CNN.	100 × 100 pixels (operating at wavelength 900–1700 mm).	750 samples.	Xiao et al. (2020)
	Foam, plastic, brick, concrete and wood.	RGB/NIR hyperspectral	Object classification with a single hidden layer forward neural network.	640 × 480 pixels (at wavelength 900–1700 mm).	250 samples for training and 150 for testing.	Xiao et al. (2019b)
	Wood, plastic, bricks, concrete, rubber and black bricks.	NIR hyperspectral	Object classification (method not specified).	Wavelength 900–1700 mm (spatial resolution not provided).	Training set of 150 samples per material type; test set includes 166 wood pieces, 130 plastic pieces, 142 red bricks, 150 concretes, 198 rubbers and 158 black bricks.	Xiao et al. (2019a)
	Concrete, brick and tile.	RGB	Object detection with SSD, YOLO and Faster R-CNN.	1920 × 1200 pixels.	4230 images for training and 1727 for testing.	Demetriou et al. (2023)
	Brick and sand-lime brick.	Line-scan camera	Object detection using SSD, YOLOv3 and Faster R-CNN.	300 × 300 pixels.	Total of 6000 images; 5000 for training, 500 for testing and 500 for validation.	Kronenwett et al. (2024)
	Bricks, concrete, plastic, metal, wood and rubber.	NIR hyperspectral, 3D camera	Object detection (specific method not detailed).	640 × 640 pixels.	2500 samples; 75% for training and 25% for testing.	Ku et al. (2021)
	Concrete bricks, plastic bottles, rubber and wood.	Hyperspectral camera, laser beam, 3D camera	Object detection using R-CNN Autoencoders.	160 × 160 pixels.	2500 grasping rectangles, with a 4:1 train:test ratio.	Wang et al. (2019)
	Concrete machine-made brick, fired brick, wood, plaster, plastic, ceramic and carton.	GB-depth	Object classification and detection via Mask R-CNN based fusion models.	Not disclosed (images scaled to 640 × 640 pixels).	3367 images; 70% for training and 30% for testing.	Li et al. (2022)
	Cardboard, soft plastic, rigid plastic and metal.	RGB	Object detection and segmentation employing RetinaNet, Mask R-CNN and TridentNet.	1920 × 1080 pixels.	10,715 images in total; 3002 for training, 929 for testing, 572 for validation and 6212 unlabelled images.	Bashkirova et al. (2022)
	Concrete bricks, tiles, foam, general waste, plasterboard, pipes, various plastics, wood and stone.	RGB	Object detection and segmentation utilizing YOLOv8.	1920 × 1200 pixels.	3127 images; 1984 for training, 573 for testing and 570 for validation.	Demetriou et al. (2024)
Laboratory setting	Concrete grains, various stones (white, grey, light coloured, slate), day bricks, ceramic tiles, bituminous grains, glass, wood, plastics, steel, paper, cardboard and other materials.	RGB	Object classification employing ResNet.	6000 × 4000 pixels (scaled to 256 × 256).	36,000 labelled database images; 2000 for training and 500 for testing in each subclass.	Lau Hiu Hoong et al. (2020)
MRF	Diverse CDW: concrete, rock, stone, bricks, various plastics such as High-Density Polyethylene (HDPE) and Low-Density Polyethylene (LDPE), metals (copper, aluminium, steel, iron), timber and general wood waste.	RGB-D	Object segmentation using an RGB-DL depth fusion strategy.	1280 × 720 pixels.	Total of 3568 images; 2500 for training, 713 for validation and 355 for testing.	Prasad and Arashpour (2024)
	Concrete, aggregates, wood, timber, hard and soft plastics, steel, cardboard, paper and mixed waste.	RGB	Combined object detection and segmentation with DuoSeg++ and DeepLab3+ Unet.	Not disclosed.	430 images; 75% for training, 15% for validation and 10% for testing.	Sirimewan et al. (2024)
Skip bin top-down view	Various timbers (second fix, shuttering/formwork), ply and particle boards, hard and soft plastics, brick, concrete, cardboard and polystyrene.	RGB	Object classification with VGG-based models.	3000 × 2250 pixels.	525 images, which includes 84 images of empty skip bins.	Davis et al. (2021)
Weigh bridge top-down view	Inert waste (concrete and bricks) and non-inert waste (wood, plastic and bamboo).	RGB	Assessing if inert waste percentage exceeds a specified level (e.g. 50%) using DenseNet169.	Not disclosed.	1127 records (comprising images and physical properties like net weight and depth), with 70% for training, 15% for validation and 15% for testing.	Chen et al. (2021)
	Rock, gravel, earth, packaging wood, other non-inert materials and mixed waste.	RGB	Object segmentation using a Boundary-aware Transformer.	1980 × 1080 pixels.	5366 images.	Dong et al. (2022)
	Rock, gravel, earth, packaging, wood, other non-inert materials and mixed waste.	RGB	Object segmentation using DeepLabv3+.	1980 × 1080 pixels.	5366 images; 3515 for training, 754 for validation and 753 for testing.	Lu and Chen (2022)

CNN: convolutional neural network; MRF: material recovery facility; CDW: construction and demolition waste; DL: deep learning.

Table 2.

A few examples of image sensing techniques (Dodampegama et al., 2024).

Sensor types	Data format	Potential features	Advantages	Disadvantages
RGB	Red, green and blue channels representing 2D spatial dimensions	Colour, size, shape, texture and brightness	Colour differentiation Cost-effective Realistic appearance Simplicity High resource availability	Lack of information Shadows and lighting effects Difficulty in differentiating material overlaps, complex wastes and surface textures
Thermal	A stack of closely packed spectral bands representing IR radiation	Heat signature, temperature differences	Heat-based detection Fewer lighting effects	Equipment costs Complex setup with preheating Complex processing Limited resource availability
RGB-D point cloud	RGB + depth in 3D spatial points	Colour, size, shape, texture, brightness and 3D spatial information	Detailed 3D visualization Accurate object localization	High equipment costs Complex processing Large data loads Limited resource availability
RGB-D	Red, green, blue and depth channels representing 2D spatial dimensions	Colour, size, shape, texture, brightness and depth	Depth information 3D visualization Surface details More realistic data	Equipment costs Complex processing Challenges in calibration: limited range and texture variation
Hyperspectral	A stack of narrow, closely packed spectral bands in visual and IR region	Spectral signatures, reflectance	Spectral differentiation Insights in chemical composition Advanced detection	Equipment costs Complex processing Large data loads Limited resource availability
Hybrid	Combination of several formats	Combination of several features	Identifying more features High accuracy High category variation	Data alignment challenges Complex processing Large data loads Limited resource availability

CV for automated waste detection and analysis

CV enables automated interpretation of images and videos, replicating human visual perception through digital image processing (Dong et al., 2022; Shahin et al., 2024). In construction, CV facilitates accurate identification of diverse waste types by analysing their physical and chemical characteristics, which influence light absorption and visible colour (Che et al., 2024; Dong et al., 2022; Lu et al., 2022; Mohsen and Martinez 2025). By extracting essential visual features, CV provides a scalable and cost-effective means of waste quantification, supporting automated sorting and reducing manual intervention. Recently, Jaiswal et al. (2025) proposed an end-to-end multi-view framework for localized CDW quantification using 3500 images. They achieved an F1 score of 0.97 for waste cluster identification and an absolute percentage error of 8.97% in volume estimation, with a processing time of approximately 11 minutes, demonstrating strong field applicability. In a similar vein, Chen et al. (2022a) created a monocular vision-based method for CDW truck payload estimate, validating it across 2914 truckloads (~800 m³/day), taking an average of 3.3 seconds for processing one image with reported relative errors of 0.065 for bucket dimension estimation and 0.169 for volume estimates. This section outlines core CV techniques, case studies and key challenges in applying CV for CDW detection and analysis.

Core CV techniques for construction waste management include object detection, image segmentation and smart collection systems. Algorithms such as Faster R-CNN and YOLO are commonly applied to identify and localize waste objects in visual data (Narayanswamy et al., 2022). Image segmentation divides images into distinct regions for material identification, converting raw pixels into structured patterns to improve classification accuracy and address occlusion issues. CNNs are widely used to extract distinguishing features for classification tasks (Narayanswamy et al., 2022; Varalakshmi et al., 2024). Smart bins equipped with sensors and AI software represent an emerging CV application that can monitor fill levels, identify waste types and detect odour thresholds in real time (Lakhouit, 2025). Comparative experiments show Faster R-CNN achieving the highest detection accuracy (91%), outperforming YOLO and CNN-based models in material recognition (Narayanswamy et al., 2022). Such intelligent systems are increasingly deployed in urban settings to optimize waste routing and resource allocation.

Numerous studies demonstrate the potential of CV for CDW analysis. Lu et al. (2022) developed a segmentation framework that identified multiple waste materials, including rock, gravel, wood and packaging, in complex outdoor environments using a dataset of 5366 labelled images. Their DeepLabv3+ model achieved high segmentation accuracy across varying site conditions. Yong et al. (2023) proposed an automated method for detecting illegal waste sites by integrating remote sensing imagery with DL models, including HRNet, DeepLabv3+ and PSPNet. The model achieved 96.3% classification accuracy and a mean IoU score of 74.6%, highlighting its reliability for large-scale landfill monitoring. Table 3 summarizes recent studies on CDW segmentation and detection, detailing materials examined, sensors used, techniques applied, and dataset characteristics across various operational environments. The table illustrates the diversity of materials, ranging from concrete and bricks to metals, plastics and wood. While RGB cameras remain the most widely used for accessibility, alternative configurations, such as RGB-D, line-scan, hyperspectral and LiDAR systems, offer enhanced feature-extraction capabilities. Advanced architectures, including YOLOv8, SSD, DuoSeg++, Swin Transformer, RACNET and Mask R-CNN, are applied across both controlled environments and dynamic on-site conditions. Despite these differences, all studies share the common objective of enhancing CDW sorting efficiency through automation, with model choice, dataset design and experimental configuration adapted to specific operational constraints and research aims.

Applying CV to real-world construction sites remains challenging due to the inherent complexity and variability of these environments. Sites are often cluttered, characterized by diverse materials, irregular layouts and constantly changing conditions, all of which hinder accurate waste quantification. Although CV technologies show strong potential to automate inspections and enable precise waste identification and quantification, their performance depends heavily on addressing constraints related to data quality, the environment and adoption.

The most critical technical limitation lies in obtaining and processing high-quality, representative data. Reliable assessment requires high-resolution images that reflect the dynamic nature of construction sites; however, frequent site reconfigurations and environmental factors such as adverse weather and visual similarity among materials degrade image quality and reduce detection accuracy (Costa et al., 2024). Models trained on controlled datasets tend to perform poorly when deployed under real conditions (Lu et al., 2022). Mixed or contaminated waste further complicates classification, as DL models typically perform best on homogeneous materials (Langley et al., 2025). Enhancing model robustness requires developing comprehensive, representative datasets that reflect the diversity of on-site waste. This entails systematic data collection, cleaning and annotation to strengthen model generalization (Rafatian et al., 2024). Beyond technical challenges, cultural and financial barriers persist. Adoption of CV-based waste management systems is often slow due to high costs, integration difficulties and uncertainty about ROI (Costa et al., 2024; Rafatian et al., 2024). Addressing these limitations requires not only improved algorithms and datasets but also effective stakeholder engagement, demonstration of cost-effectiveness and seamless integration of CV solutions into existing construction workflows.

AI and ML for automated waste quantification

The adoption of AI through advanced analytical and sensing tools is redefining modern construction waste management. Beyond improving design efficiency and minimizing environmental impact (Regona et al., 2024), AI-driven pattern recognition and predictive analytics enhance resource use and streamline waste handling. Smart collection systems equipped with sensors and AI software address inefficiencies in manual operations, such as irregular schedules and poor routing, through real-time monitoring and adaptive route planning. DL models can classify materials based on visual or chemical properties, forecast waste generation trends and integrate with robotic systems for automated sorting. The increasing deployment of robotics for on-site and off-site CDW sorting enhances efficiency, accuracy and worker safety. Collectively, these innovations are reshaping CDW management practices and accelerating the sector’s transition towards CE goals (Chen et al., 2022b; Dodampegama et al., 2024; Olawade et al., 2024).

Traditional ML algorithms, such as RF, support-vector machines (SVMs), gradient boosting, logistic regression, ANN and decision trees, can classify and detect waste items, but often struggle to differentiate fine-grained material variations within heterogeneous waste streams (Midigudla et al., 2025). Recent AI-driven image-classification and object-detection models, leveraging CNNs and other advanced architectures, are increasingly used to automate both waste categorization and volumetric estimation (Abdu and Mohd Noor, 2022). These tools streamline data collection, enhance operational efficiency and support sustainable resource recovery (Xia et al., 2022). Few studies, however, have tested DL performance on mixed or contaminated wastes under real-world material-recovery conditions (Langley et al., 2025). Lin et al. (2022) employed knowledge-transfer strategies to optimize DL for CDW classification, while (Cha et al., 2022) proposed a hybrid model combining SVM regression with ANN architectures to overcome small-dataset limitations. In contrast to traditional ML techniques, DL models, particularly CNNs, can automatically learn hierarchical feature representations, making them more effective for complex CV applications in waste sorting (Kang et al., 2022).

Deep-learning approaches and model training

DL has advanced rapidly and now underpins most automation in CDW quantification. CNNs are highly effective for image-based waste sorting, enabling the classification of materials into categories such as general, organic and recyclable waste. Models such as ResNet50, GoogleNet and InceptionV3 are frequently used to benchmark accuracy, while sensor-linked DL frameworks help optimize waste-collection routes and reduce operational costs (Xia et al., 2022). Both ANNs and CNNs have proven useful for auditing waste data: ANNs capture complex, nonlinear relationships but require extensive data and remain difficult to interpret, whereas CNNs efficiently process visual data for sorting and quantification (Samal et al., 2025). Recently, Al-Mashhadani (2023) analysed and evaluated various DL models used for waste materials sorting and classification.

Model training typically relies on datasets collected from construction sites that capture different debris types and volumes (Lin et al., 2022; Na et al., 2022). Proper validation is essential to ensure generalizability, yet many studies overlook this requirement, particularly when sample sizes are limited. No universal benchmark exists for selecting the best algorithm for CDW classification (Gao et al., 2024). Nevertheless, CNNs, recurrent neural networks and deep neural networks remain the most widely used architectures for computer-vision-based waste recognition, as they can analyse multidimensional data and automatically extract hierarchical features (Farghali and Osman, 2024).

Success stories, challenges and limitations

Several case studies demonstrate AI’s growing role in CDW management. Jiang et al. (2022) developed a cost-effective system integrating drone scanning, DL and Geographic Information Systems (GIS) to track CDW at project sites and unauthorized dumps. Their approach used 3D reconstruction via structure-from-motion, image segmentation with fully convolutional networks and GIS-based visualization for spatial monitoring. Farshadfar et al. (2025) compared manual and ML-based sorting technologies in Finland, identifying differences in cost, environmental performance and operational efficiency. Nežerka et al. (2024) achieved classification accuracies of 85.9% using CNN, 92.3% using Gradient Boosted Decision Tree (GBDT), and 91.3% using Multilayer Perceptron (MLP) for debris fragments captured with RGB cameras, as illustrated in Figure 9. Another recent study by Cheng et al. (2024) applied the YOLOv5 model to UAV imagery, achieving precisions of 0.88 for concrete and 0.96 for brick.

Figure 9.

Identification of image-based recognition of CDW using RGB camera with ML models (reproduced from Nežerka et al. (2024) with permission, License number: 6178901419943).

CDW: construction and demolition waste; ML: machine learning.

Despite notable progress, several obstacles persist. High-quality, standardized datasets remain scarce, limiting model accuracy and generalization (Farghali and Osman, 2024; Olawade et al., 2024). Implementation requires substantial investment in hardware, software and skilled personnel, creating cost barriers. The heterogeneity of CDW complicates classification and model transferability, while algorithm selection remains non-trivial. Many systems still focus on classification rather than volumetric or mass estimation, reducing their logistical usefulness. A lack of standardized evaluation metrics, such as IoU, also hinders benchmarking and cross-study comparison. Addressing these challenges will enable AI- and ML-based frameworks to deliver measurable improvements in decision-making, efficiency and sustainability in construction waste management.

Integrating automation technologies for real-time waste management

Real-time management of CDW increasingly relies on automation technologies for measurement, inspection and process optimization. Maintaining productivity and sustainability requires adaptive digital systems capable of streamlining waste handling (Chen et al., 2022b; Patil et al., 2024). Modern solutions, such as BIM, IoT and industrialized building systems, have substantially reduced waste generation through early forecasting, remote sensing and real-time data integration. IoT systems enhance detection by combining sensor data with DL to identify waste types (Afzaal and Ul Haq, 2025), while IoT-enabled sorting machines track material composition in real time (Lakhouit, 2025). For instance, Azath et al. (2025) applied IoT with gradient boosting to optimize waste and resource flows through sustainable industrial symbiosis. Automated systems have demonstrated potential to reduce waste output, lower operational costs and improve site efficiency, although high set-up costs and limited interoperability still constrain large-scale adoption (Abkar et al., 2023).

Digital integration of BIM, IoT and real-time monitoring

Effective waste management on dynamic construction sites depends on continuous monitoring and accurate quantification. Real-time data allows precise waste prediction, regulatory compliance and faster corrective actions. Predictive models based on linear regression have achieved strong correlations (adjusted R² = 0.877–0.893) in forecasting residential waste generation, with key variables including site organization, design uniformity and total project area (KhairEldin et al., 2025).

BIM has emerged as a core digital tool for material quantification, enabling simulations that optimize design decisions, reduce waste and prevent rework. The construction industry, responsible for nearly half of global solid waste, has reported measurable reductions through BIM-assisted workflows. South Korean case studies recorded waste reductions of 4.3–15.2%, while one project achieved a 56% decrease using BIM-integrated estimation (Aftab et al., 2024). Despite these benefits, broader adoption is limited by data collection gaps, software incompatibility and interoperability issues (Han et al., 2021; Olubambi et al., 2024). IoT technologies complement BIM by providing real-time data on material usage and waste generation (Jadoon et al., 2025; Lakhouit, 2025). Sensors embedded in construction equipment and bins monitor material consumption and waste accumulation, enabling timely adjustments (Lopes et al., 2024). One notable study utilizes a Smart BIM framework that integrates IoT to manage demolition waste, enabling practical quantification, route optimization and data-driven disposal planning. The system improved sustainability outcomes and enhanced decision-making among project managers (Kang et al., 2022). These platforms strengthen sustainability performance by linking on-site conditions to digital project models.

Several other IoT-based innovations highlight the progress in digital waste quantification. Anh Khoa et al. (2020) developed an IoT–ML framework for predicting waste levels and optimizing collection routes, demonstrating real-time gains in waste-management efficiency. Pujari et al. (2024) integrated IoT with ML to enhance construction-site safety, indirectly minimizing waste through better process control. Chandrasekaran et al. (2023) used IoT and GIS to monitor CDW dump yards, improving policy enforcement and environmental oversight. Belhiah et al. (2024) proposed an IoT-enabled urban waste analytics platform that optimized debris collection and reduced ecological impacts. Collectively, these studies show how IoT and BIM integration can transform CDW management by merging monitoring, analytics and operational control into a unified digital ecosystem.

AI, CV and automated feedback systems

Recent advances in digital imaging, CV and AI have revolutionized continuous waste tracking. Vision-equipped drones and AI-based monitoring systems can localize and classify waste in real time, particularly across large or complex construction sites. Integrated AI–IoT platforms provide operational insights and enhance sorting efficiency through automated bin measurement and feedback analysis. These systems continuously retrain models with new sensor data, improving accuracy over time. Vision-based detection technologies now identify material types such as plastics, metals and aggregates using spectral and visual signatures, enhancing sorting precision and recyclability (Recycleye, 2025; Singh, 2023; Viso.ai, 2025; Wang et al., 2024). The convergence of AI, IoT and BIM has also enabled automated feedback loops that continuously refine construction-site performance. Studies confirm that automation technologies increase productivity and reduce material waste (Abkar et al., 2023). Feedback mechanisms, both positive and negative, play a critical role: Positive loops reinforce effective practices such as prefabrication and on-site recycling, while negative loops highlight inefficiencies when waste generation rises despite new technologies (Li et al., 2014). Moreover, some researchers have demonstrated that feedback loops modelled using system dynamics can simulate relationships among waste generation factors, allowing stakeholders to test alternative strategies virtually and make data-driven, goal-aligned decisions (Mahinkanda et al., 2023).

Overall, because of their scalability, automation potential and compatibility with digital construction workflows, AI-driven robotic sorting, UAV-based monitoring and IoT-integrated sensing platforms currently show the greatest potential for real-world deployment. On the other hand, even though they are innovative, exclusively imaging-based or isolated experimental prototypes are still restricted by the scarcity of datasets, their sensitivity to diverse waste streams and the expense of implementation. Therefore, future research should focus on increasing system robustness, reducing hardware and operating costs, testing performance across various field conditions and improving compatibility with BIM and sensor networks. These initiatives will close the gap between laboratory innovation and operational adoption by coordinating methodological progress with practical needs, offering a roadmap for the upcoming generation of automated CDW management systems.

Challenges and limitations

Automated waste quantification systems face several challenges that hinder their effectiveness and large-scale implementation, including data limitations, algorithmic constraints and limited adaptability to diverse material types. Overcoming these issues requires algorithms capable of handling complex, real-world scenarios and adaptable frameworks that evolve in response to industry needs, while remaining accessible to all stakeholders. Table 4 synthesizes the key barriers and limitations of the technologies reviewed in this study. These limitations are evident across several key areas:

Table 4.

Barriers across automated waste-sorting technologies.

Issue	AI/ML methods	BIM approaches	IoT/sensors	Imaging/CV	UAV systems
Data readiness	Limited site-diverse training data	Waste data often missing	Short-term, sparse datasets	High labelling effort	Few standardized aerial datasets
Field robustness	Weak generalization across sites	Model quality dependent	Noise and drift issues	Sensitive to dust and lighting	View-angle and motion errors
Interoperability	Outputs poorly aligned with tools	Limited real-time linkage	Multi-sensor fusion challenges	Non-unified data formats	Weak BIM/database integration
Cost	High compute and development cost	Model creation/updates costly	Hardware and maintenance cost	High-resolution hardware cost	Purchase, operation, licensing
Hardware limits	GPU/edge constraints	Storage and computing load	Power and durability limits	Fixed placement constraints	Battery and weather limits
Deployment gap	Mostly pilot-scale validation	Design-stage dominance	Limited long-term field use	Controlled-environment bias	Regulatory and site constraints
Research need	Transferable, explainable models	Dynamic BIM updating	Robust long-term deployment	Site-tolerant vision models	Repeatable operational workflows

AI: artificial intelligence; CV: computer vision; BIM: building information modelling; IoT: Internet of things; UAV: unmanned aerial vehicle; ML: machine learning; GPU = Graphics Processing Unit.

1. Data quality remains one of the most critical determinants of automated waste classification accuracy. Many training datasets used for ML are small, narrowly scoped and fail to reflect the complexity of real-world waste streams. Public datasets often lack diversity in shape, colour and texture, while inconsistent labelling reduces model accuracy and generalization, leading to unreliable predictions.

2. Algorithmic performance is another constraint. Even advanced architectures, such as CNNs, can experience significant performance declines in dynamic, unstructured environments. Lighting variability, occlusions and overlapping materials frequently impair classification accuracy.

3. Adaptability across contexts also remains limited. Models trained on localized datasets often perform well under specific conditions but fail when applied to regions with different waste compositions. More comprehensive, representative datasets and improved interoperability with existing municipal systems are needed, although integration remains technically complex and financially demanding.

4. High initial and operational costs continue to pose barriers, particularly in regions with low investment in waste infrastructure. Although automated technologies for sorting CDW, such as robotic arms, CV systems and sensor-based separation platforms, offer considerable productivity benefits, actual applications frequently encounter difficulties with data compatibility, cost and system integration. For instance, massive datasets produced by UAV, RGB and hyperspectral imaging systems often generate large volumes of unstructured visual data that are not directly compatible with BIM platforms or waste-inventory databases and require preprocessing before integration with management software, which increases implementation time and complexity. Adoption of high-end hardware, such as robotic sorting units, LiDAR scanners, UAV platforms, high-resolution cameras or AI-enabled conveyor systems, may be restricted to large-scale facilities or pilot projects due to their high initial costs. Furthermore, as various case studies have shown, integrating automated sorting systems into existing waste processing workflows often requires software adaptation and skilled staff, increasing operational complexity and delaying project execution. These findings highlight the importance of interoperability, hardware cost, integration design and algorithmic correctness in scaling automation for construction waste sorting.

5. The shortage of skilled personnel further constrains implementation. Effective operation demands technically trained staff to manage, maintain and calibrate systems. The absence of such expertise often necessitates third-party support, increasing costs and dependency.

6. Additional challenges include data privacy, liability and ethical concerns. Automated systems collect sensitive information on construction processes and worker activities, requiring strict adherence to data protection laws such as the European General Data Protection Regulation (GDPR). Security breaches can lead to legal and financial repercussions, while concerns about surveillance and inadequate communication can erode worker trust and hinder informed consent. Ethical deployment of automated monitoring technologies, including imaging systems, embedded sensors or UAV-based data collecting, necessitates adherence to institutional rules, data protection legislation and relevant local regulations. A minimal level of privacy protection must be ensured in practical implementations by avoiding personally identifiable information, anonymizing recorded images and limiting data use to structural or material evaluation purposes only.

To overcome these challenges, future research should focus on improving data quality, algorithmic robustness and system integration within urban infrastructure to support large-scale deployment. Priorities include developing diverse, publicly available datasets; advancing hybrid algorithms that combine multiple modelling techniques for higher accuracy; and establishing unified data standards and protocols to enable interoperability among stakeholders. To address data scarcity, synthetic data generation and digital twinning can be used to simulate various CDW generation scenarios, thereby augmenting limited field data. Federated learning frameworks facilitate citywide knowledge generation by enabling cooperative training of AI models across multiple locations without exchanging sensitive raw data. In parallel, IoT-enabled urban sensor networks, supported by standardized protocols and cloud-based data environments, can facilitate near real-time CDW data capture, aggregation and sharing across stakeholders at scale. BIM–IoT interoperability can be strengthened through open standards, semantic alignment and integration architectures. Open BIM formats (e.g. IFC) and clear exchange schemas (e.g. JSON-based sensor mappings) can translate time-stamped IoT streams into BIM-compatible data. Reliable linkage between sensor observations and BIM objects can be achieved using unique object identifiers and standardized metadata (e.g. location, timestamp, sensor type and units). Middleware (e.g. API gateways, message brokers and ETL pipelines) can then filter, validate and transform heterogeneous IoT data before integrating it into a common data environment or a digital twin. Cloud-based integration services further support scalability through authentication, version control and event-driven updates, while reducing conflicts between proprietary BIM platforms and vendor-specific IoT systems. Moreover, research on user interaction and long-term system performance is vital for ensuring usability, transparency and reliability. Addressing these challenges through better data, smarter algorithms and standardized frameworks will accelerate the transition towards scalable, city-wide automated waste quantification.

Conclusions

Accurate and timely quantification of CDW is a pivotal step towards achieving CE goals in the construction sector. While conventional methods such as per capita multipliers, generation-rate models and site audits provide baseline estimates, they often lack the material-level detail, real-time capability and adaptability required for modern waste streams. Recent advances in digital technologies, including BIM, IoT-based sensing, UAV photogrammetry, hyperspectral imaging and AI-driven CV, are enabling more precise and auditable measurements, allowing waste flows to be tracked from generation to recovery with unprecedented accuracy. This review synthesizes global progress in CDW quantification, identifies persistent gaps between high waste generation and low recovery performance and evaluates emerging AI- and sensor-enabled methods for real-time, material-level estimation. It also proposes a practical roadmap for integrating these technologies into scalable, interoperable and standardized quantification systems. Embedding such systems into operational workflows can transform waste data into a decision-ready resource across all phases of a project, thereby strengthening both policy formulation and on-site management practices. As illustrated by global disparities, such as China’s high CDW generation but low recovery rates, the absence of integrated, real-time measurement frameworks can undermine even large-scale waste initiatives. Moving forward, research and industry efforts should prioritize cross-platform interoperability, robust validation under real-world conditions and the establishment of common performance benchmarks. By doing so, AI-enabled quantification can progress from experimental demonstrations to an operational standard that supports measurable advances in waste prevention, material recovery and circular construction.