Estimation, Classification, and Prediction of Construction and Demolition Waste Using Machine Learning for Sustainable Waste Management: A Critical Review

Abstract

The management of construction and demolition waste is a critical concern for sustainable urban development and environmental conservation. In this review, the authors provides an overview of the involvement of machine learning techniques like the support vector machine (SVM), artificial neural networks (ANNs), Random Forest (RF), K-nearest neighbor (KNN), deep convolutional neural networks (DCNNs), etc. in the estimation, classification, and prediction of construction and demolition waste, contributing to the advancement of sustainable waste management practices. The authors observed that the DCNN achieved an outstanding accuracy of 94% in the estimation and classification of construction waste. Based on the authors’ observations, the machine learning models are well suited for the prediction or classification of construction waste and are good for sustainable waste management in the future. This paper provides insights into the promising future of machine learning in revolutionizing waste management practices and future research.

1. Introduction

Construction and demolition (C&D) waste represents a significant environmental challenge, comprising debris generated during the construction, renovation, and demolition of buildings [1,2,3]. Defined under India’s Construction and Demolition Waste Management Rules [4], C&D waste includes both inert and non-inert materials, such as concrete, steel, glass, wood, and rubble [4,5]. Globally, C&D waste contributes substantially to landfill congestion, pollution, and resource depletion, with its management being pivotal for sustainable development [6]. CDW accounts for up to 40% of the total waste generated in the European Union and close to 67% of that in the United States [7]. A pie chart illustrating the bulk percentage contribution of garbage categories dumped at Askar Landfill from 1997 to 2016 is shown in Figure 1. The most common types of garbage were household (31%), construction and demolition (27%), and business trash (23%). The other waste types account for 19% of all waste that is disposed of.

The rapid growth in global waste generation poses a significant environmental and public health challenge that requires urgent attention. According to projections, by 2050, global waste generation is expected to rise by 70%, reaching 3.40 billion tons annually from 2.01 billion tons in 2016 [8,9]. C&D waste is the third-largest contributor among special waste categories, highlighting the urgent need for efficient management strategies [10].

The construction industry is among the most resource-intensive sectors worldwide, responsible for 40% of global energy consumption and 16% of total water usage. It also accounts for the extraction of approximately 40% of the Earth’s raw materials and 25% of virgin wood for its activities [6]. With growing resource scarcity, C&D waste recycling has become critical. Currently, only 12% of materials used in construction come from recycled sources [11]. Furthermore, the use of industrial waste materials such as fly ash in the construction lifecycle presents a tremendous opportunity for the industry, contributing to sustainability and cost-effectiveness while improving energy performance [12]. However, aging infrastructure is expected to yield 270 billion tons of demolition material globally over the next two decades, offering vast recycling potential [13]. C&D waste management is central to achieving a circular economy, reducing landfill dependency, and addressing the economic burden of material wastage. Approximately 21–30% of cost overruns in construction projects are attributed to material wastage [14]. C&D waste management is fraught with challenges, such as the accurate estimation, classification, and prediction of waste generation [15,16,17]. Minimizing rebar and concrete use cuts costs, reduces waste, lowers emissions, and improves resource efficiency throughout construction [18].

Regional-level data on waste generation are crucial for developing effective waste management strategies, yet many regions, especially emerging ones, lack reliable information of this nature [19]. Iyiola et al. [20] identifies seven key Digital Technologies—blockchain, the IoT, AI, ML, Robotics, CV, and BIM—that are crucial for enhancing C&DWM efficiency in the construction sector, highlighting their potential to guide stakeholders and policymakers towards overcoming adoption barriers and fostering more sustainable C&DWM practices.

These difficulties stem from diverse material compositions, contamination, and inadequate data. Advanced tools like machine learning (ML) have shown potential in addressing these issues by analyzing large datasets, improving waste categorization, and predicting trends [19,21].

Machine learning (ML) techniques offer transformative potential in optimizing waste management [22]. ML can improve waste estimation [23], enable accurate classification, and predict trends to aid resource allocation and recycling. By integrating ML technologies, the construction sector can enhance sustainability, reduce costs, and address environmental concerns. Machine Learning (ML) has emerged as a powerful tool in waste management due to its ability to process large datasets, identify patterns, and make accurate predictions.

Machine learning can predict future waste volumes by analyzing historical data on generation rates, population growth, and economic factors, while also estimating waste composition to support recycling and recovery efforts [21,24]. Machine learning algorithms can classify waste items through image analysis for automated sorting and can analyze sensor data from containers or landfills to identify waste types and contamination [25]. Machine learning can predict equipment failures by analyzing sensor data for preventive maintenance and can forecast future waste trends through historical data analysis, aiding proactive planning and resource allocation [26].

The primary objective of this review paper is to provide a comprehensive overview of the existing research and methodologies employed in estimating, classifying, and predicting construction and demolition (C&D) waste using machine learning techniques. The paper aims to (a) identify and analyze the various machine learning algorithms and techniques that have been applied to C&D waste management; (b) evaluate the effectiveness and limitations of these methods in addressing specific challenges, such as waste estimation, classification, and prediction; (c) highlight the potential benefits of machine learning in improving the sustainability and efficiency of C&D waste management practices; and (d) identify research gaps and future research directions to advance the field and enhance the application of machine learning in C&D waste management.

2. Review Methodology

This study employed a structured and systematic review methodology to explore the application of machine learning in the estimation, classification, and prediction of C&D waste. The paper is organized into four sections: systematic review study selection, findings and discussions on three key aspects, critical insights, and potential directions for future research (Figure 2). Bibliometric analysis was conducted using VOSviewer ver. 1.6.20 software, examining parameters such as the evolution of publications over time, publication sources, international collaboration patterns, and co-occurrence analysis to identify research hotspots. These analyses provided a comprehensive statistical mapping of bibliometric data related to scientific publications on C&D waste.

2.1. Systematic Review Study Selection

For this study, the international databases Scopus and Web of Science were selected as primary sources due to their comprehensive coverage and user-friendly interfaces. These multidisciplinary databases include peer-reviewed journal articles, books, and conference proceedings, making them ideal for academic research. The search keywords used were “(CONSTRUCTION OR DEMOLITION) AND WASTE AND (ESTIMATION OR CLASSIFICATION OR PREDICTION OR (MACHINE AND LEARNING) OR AI OR (ARTIFICIAL AND INTELLIGENCE) OR ANN OR (ARTIFICIAL AND NEURAL NETWORK) OR IOT OR (INTERNET OF THINGS))”. The search yielded a total of 5674 articles (1760 from Web of Science and 3914 from Scopus).

Initial screening was performed using three inclusion criteria: (1) articles in subject areas such as Engineering and Computer Science, (2) publications in English, and (3) studies published within the last ten years (2015–2025) to focus on recent advancements. This process narrowed the selection to 1996 articles, from which 370 duplicates were removed, leaving 1626 articles for further analysis. These were subsequently screened through a two-step review process, (Step 1) title and abstract review, and (Step 2) full-text review, ultimately resulting in 58 articles for in-depth evaluation.

Sixty-seven papers in all were selected for critical debate and both qualitative and quantitative analysis. Figure 3 summarizes the study selection strategy. The titles and abstracts of the 61 chosen papers were analyzed using text-mining techniques and a co-occurrence algorithm in VOSviewer. When paired with keyword co-occurrence analysis, this approach yielded insightful information on the field’s conceptual frameworks and main study topics.

2.2. Annual Distribution of Publications

The analysis of publications on the application of machine learning for the classification, estimation, and prediction of construction and demolition (C&D) waste shows a significant growth trajectory over the years and is shown in Figure 4. Starting with minimal contributions in 2012 (2 publications) and 2013 (3 publications), research activity remained relatively low until 2019, with most years having only 1–2 publications. However, a noticeable surge began in 2020, which saw 7 publications, reflecting an increased interest in leveraging machine learning techniques for addressing C&D waste challenges. This upward trend continued, peaking in 2023 with 14 publications, followed closely by 12 publications in 2024. The substantial rise in research output from 2020 onwards underscores the growing importance of machine learning in this domain, driven by advancements in technology and an increasing focus on sustainable construction practices.

2.3. Publication Type

This compilation of research contributions includes a diverse range of scholarly outputs. It comprises 3 books/acts/rules, demonstrating a foundational contribution to the field; 7 conference papers, reflecting active participation and the dissemination of findings within academic and professional communities; and 56 articles, highlighting extensive original research published in peer-reviewed journals (Figure 5).

2.4. Source Analysis

This study reviews publications focusing on the application of machine learning in the classification, estimation, prediction, and management of construction and demolition (C&D) waste (Figure 6). The analysis highlights contributions across various reputable journals. Waste Management leads with 5 publications, followed by Resources, Conservation and Recycling and Sustainability (Switzerland), each contributing 4 publications. Other significant sources include the Asian Journal of Civil Engineering, the International Journal of Environmental Research and Public Health, the Journal of Cleaner Production, and Waste Management and Research, each with 3 publications. Additionally, Automation in Construction and Buildings contribute 2 publications each, while Construction and Building Materials accounts for 1 publication. This distribution reflects the interdisciplinary interest and the growing relevance of machine learning in addressing challenges associated with C&D waste.

2.5. Publication Distribution by Countries

The distribution of publications by country was analyzed to assess the countries’ contributions and collaborations in the development of research on machine learning for construction and demolition waste estimation, classification and management using VOSviewer software. Citations and co-authorship were key metrics used in the analysis. Out of 32 countries considered based on the search criteria, only 11 met these criteria.

Table 1. Productive countries engaged in C&D research using machine learning.

Country	Documents	Citations	Total Link Strength
United Kingdom	4	369	5
United States	5	593	5
China	10	413	4
Australia	5	403	3
Iran	2	99	2
Hongkong	3	92	1
India	9	65	1
Malaysia	2	24	1
Brazil	4	315	0
South Korea	8	94	0
Turkey	2	41	0

details these 11 productive countries, their number of articles, citations, and total link strength. Figure 7 represents the contributions of 11 countries, where larger circles indicate higher publication outputs and thicker lines signify stronger collaborative ties between countries. This visualization highlights countries with significant contributions and robust collaborations in the field. The findings underscore the global relevance of addressing C&D waste using machine learning approaches.

2.6. Co-Occurrence Analysis of Keywords

The co-occurrence analysis of keywords provides critical insights into the thematic structure and research focus in the field of the classification, estimation, prediction and management of construction and demolition waste. This analysis, performed using VOSviewer software, explores the interconnections between keywords in academic publications, helping to identify dominant topics and emerging trends. A total of 618 keywords were analyzed, with a threshold set for keywords occurring at least 5 times. Out of the total, 30 keywords met this criterion. The analysis identified 30 clusters, 369 links, and a total link strength of 1126, indicating robust interconnections between various research themes. Figure 8 highlights the interrelationship between keywords based on their co-occurrence. Dense connections around “waste management” and “machine learning” highlight these as central research areas. Strong links between “construction waste”, “demolition”, and “recycling” emphasize the importance of sustainability in the field. Terms like “machine learning”, “artificial intelligence”, “artificial neural network”, and “deep learning” demonstrate the increasing application of machine learning in C&D management.

Figure 9 provides a density visualization, indicating the frequency and importance of keywords. The density visualization is a heatmap-representation that identifies keywords, using the software VOSviewer, such as “waste management”, “machine learning”, and “construction and demolition waste” as focal points due to their high occurrence.

The keyword “waste management” appears as a central node with strong linkages to other terms, reflecting its pivotal role in the research field. Keywords like “construction and demolition waste” and “demolition waste” are prominently linked with terms such as “recycling” and “sustainability”, signifying the focus on sustainable waste handling.

3. Traditional Method of Estimation, Classification, and Prediction of CDW

3.1. Estimation

Effective waste management in construction and demolition relies on accurately estimating waste volume and composition through traditional methods like manual surveys and waste characterization, as well as modern techniques such as Building Information Modeling (BIM) and advanced software tools [27]. Key techniques include assessing on-site waste, analyzing waste types, and using the Bill of Quantities (BoQ) method to predict waste based on materials.

3.2. Classification

Visual inspection categorizes waste by observable traits, while material analysis uses lab techniques for detailed assessments. Other methods include weight and volume measurement and sorting trials [28]. Challenges like mixed waste and regulatory compliance complicate classification. However, combining traditional methods with technology-assisted sorting can improve accuracy and efficiency in managing construction and demolition waste [29].

3.3. Prediction

Traditional methods for predicting construction and demolition (C&D) waste generation include statistical models like regression analysis and time series analysis, as well as expert judgment through the Delphi method and historical benchmarking [30,31]. Waste generation rates per unit of construction also aid estimates. Challenges include data limitations, variability in waste generation, and changing regulations. While they are valuable, these methods could benefit from modern techniques like machine learning to improve accuracy and adaptability [22].

3.4. Results and Discussions

Table 2. Summary of machine learning models used for prediction of C&D waste.

Author(s)	Model(s) Used
[1]	ANN
[2]	Multi-layer perceptron artificial neural network (MLP-ANN)
[3]	Monocular vision, Deep learning model
[5]	Single-stage (SSD, YOLO) and two-stage (Faster-RCNN) detectors with various backbone feature extractors (ResNet, MobileNetV2, EfficientDet)
[15]	Hybrid AI models using CATPCA, ANN (MLP), SVMR, CATPCA–ANN (MLP), and CATPCA–SVMR
[16]	Deep learning framework in R programming: DNN, RNN, CNN
[17]	ANN (neural networks with two, five, or ten neurons in the hidden layer)
[20]	CAD, Linear regression models
[22]	Decision tree and KNN, Neural Network
[23]	Logistic regression, kernel SVM, KNN, RF, XGBoost, CatBoost
[24]	Various ML algorithms: ANN, KNN, LR, RF, and SVM
[25]	CNN
[27]	BIM, IoT
[31]	Decision tree and KNN
[32,33]	Principal Component Analysis (PCA) with decision tree, K-nearest neighbors, and linear regression algorithms
[34,35]	Decision tree (DT)-based ensemble models: Random Forest—RF, extremely randomized trees—ET, Gradient Boosting Machine—GBM, and extreme gradient boost—XGboost
[36]	AI (general term)
[37]	CNN-LSTM SWOT Analysis
[38]	LR, Pre-Renovation Auditing PRA, machine learning regression
[39]	Support vector machine regression (SVMR) Model, Categorical Principal Components Analysis (CATPCA)
[40]	Random Forest (RF) and Gradient Boosting Machine (GBM)
[40]	Random Forest (RF)
[41]	ANNs (neural networks with one and two hidden layers)
[42] [43]	GWO-ANN, AOA-ANN ANN

summarizes machine learning models employed for predicting waste generation in construction and demolition projects. The models range from classical machine learning approaches like decision tree (DT), K-nearest neighbors (KNN), and linear regression (LR), to advanced ensemble methods such as Random Forest (RF), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost). Neural networks, including multi-layer perceptrons (MLP-ANN), convolutional neural networks (CNNs), and hybrid AI models like CNN-LSTM, are also widely applied. Additionally, feature extraction techniques such as Principal Component Analysis (PCA) and Categorical Principal Component Analysis (CATPCA) are used in combination with AI algorithms.

Emerging technologies, including the Internet of Things (IoT), Building Information Modeling (BIM), and pre-renovation auditing (PRA), are integrated into waste prediction workflows. Deep learning frameworks like YOLO, SSD, and Faster-RCNN leverage advanced backbone feature extractors (ResNet, MobileNetV2, and EfficientDet). Other studies explore optimization algorithms such as Grey the Wolf Optimizer (GWO) and the Arithmetic Optimization Algorithm (AOA) with neural networks.

Table 2 highlights a growing trend towards using hybrid, ensemble, and deep learning models for more accurate and efficient predictions in construction waste management.

4. Review of Machine Learning Applications for Estimating C&D Waste

Various machine learning (ML) models have emerged as powerful tools for improving the accuracy and efficiency of C&D waste estimation by leveraging historical data, sensor inputs, and real-time project parameters such as regression models, deep learning techniques, and hybrid techniques.

4.1. Regression Models

Regression models for predicting construction and demolition (C&D) waste include linear regression, which is simple and effective for smaller projects but assumes linear relationships [44], nonlinear regression, which captures complex relationships for large-scale projects, and ensemble models like Random Forest and Gradient Boosting [41], which combine multiple models for improved accuracy but are more complex and computationally intensive.

4.2. Deep Learning Techniques for Estimation

Deep learning techniques for predicting construction and demolition (C&D) waste include artificial neural networks (ANNs), which model complex, nonlinear relationships using historical data but require large datasets and are difficult to interpret, and convolutional neural networks (CNNs) [15,25], which analyze images for waste sorting and characterization, offering high accuracy but also needing extensive data and being computationally intensive. Figure 10 shows some of the network architectures utilized in ANNs. Figure 10a shows the 2-2 network, which has two neurons in its output layer and one hidden layer with two neurons, supporting two outputs. Figure 10b, on the other hand, shows the 2-2-2 network, which has two hidden layers with two neurons each and an output layer with two outputs [42].

4.3. Hybrid Approaches

Hybrid approaches in construction and demolition (C&D) waste management combine traditional methods, like manual surveys, with machine learning [39] to enhance waste estimation accuracy using historical data and real-time inputs from sensors. These systems leverage expert knowledge and data-driven insights to create adaptive waste management plans but require complex integration and can be costly to develop and maintain [16].

Machine learning models, particularly regression, deep learning, and hybrid approaches, have the potential to transform how C&D waste is estimated and managed. They provide better accuracy, automation, and scalability compared to traditional methods like manual surveys and waste characterization studies. However, the success of ML applications depends on the availability of large datasets, computational power, and the integration of these models with existing waste management processes.

5. Review of Machine Learning Applications’ Classification of C&D Waste

Machine learning applications for classifying construction and demolition (C&D) waste are increasingly gaining traction, leveraging various techniques to enhance accuracy and efficiency in waste management. Here is a review of key areas:

5.1. Image Processing and Computer Vision Techniques

Image processing and computer vision techniques for construction and demolition (C&D) waste management include object detection and image segmentation. Object detection, using frameworks like YOLO [5] and Faster R-CNN, enables the real-time recognition of materials such as concrete, wood, and metal, facilitating automatic sorting and increasing efficiency. Image segmentation, utilizing methods like U-Net [45] and Mask R-CNN, divides images into segments for detailed analysis, offering insights into the composition and quantity of materials in mixed waste streams.

5.2. Feature Extraction and Classification Algorithms

Feature extraction and classification algorithms, such as support vector machines (SVMs) and Random Forests [46], are vital for waste classification in construction and demolition (C&D) management. SVMs find the optimal hyperplane to separate different waste classes using features from images or sensor data, which is effective for small to medium datasets but less so for larger, complex ones. Random Forests enhance classification accuracy by combining multiple decision trees, making them robust against overfitting and suitable for large feature sets, though they can be computationally intensive.

5.3. Deep Learning Techniques

Deep learning techniques like convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are effective for waste classification in construction and demolition (C&D) management [47,48]. Demetriou et al. [5] assessed deep learning models for CDW localization and classification. They tested 18 models using the first public CDW dataset, finding that YoloV7 achieves 70% accuracy and the fastest inference (<30 ms). Faster-RCNN models exhibited less mAP fluctuation, except for YoloV7, which excelled. Faster R-CNN ensures a stable mAP through two-stage detection and refined feature extraction, enhancing precision in C&D waste detection. YOLOv7’s single-stage approach prioritizes speed but introduces variability due to anchor mechanisms and sensitivity to small, overlapping objects.

Research by Davis et al. [47] and Dantas et al. [1] focused on classifying and forecasting the properties of specific waste materials using an ANN and a DCNN, achieving significant accuracy levels for compressive strength and mixed waste classification. CNNs excel at processing images, automatically extracting features and achieving high classification accuracy, minimizing the need for manual feature engineering [49]. RNNs, while primarily used for sequential data, can analyze temporal features in video data of waste sorting processes, enhancing monitoring and classification over time. However, RNNs may demand more computational resources and complex architectures than CNNs.

The integration of machine learning techniques, particularly in image processing and deep learning, holds significant potential for improving the classification of C&D waste. By combining various methods, such as CNNs for image classification with traditional classifiers like SVM and Random Forests for feature-based classification, waste management systems can achieve higher efficiency and accuracy. As technology advances, these methods are likely to become more refined, leading to more sustainable waste management practices.

6. Review of Machine Learning Applications’ Prediction of C&D Waste

Machine learning applications for predicting construction and demolition (C&D) waste generation have evolved significantly, with methods like ARIMA and LSTM playing key roles in time series analysis and forecasting. Here is a review of these models:

6.1. Time Series Analysis and Forecasting Models

A.

ARIMA (Auto Regressive Integrated Moving Average) is a time series-forecasting technique that combines autoregressive, integrated, and moving average components [46]. It is used to analyze historical waste generation data and predict future quantities, offering interpretable results for univariate series with clear patterns. However, it assumes linear relationships and requires stationary data, often needing preprocessing like differencing [50].

B.

A LSTM (Long Short-Term Memory) Network is a type of recurrent neural network designed to learn from sequential data and capture long-term dependencies. It is particularly useful for modeling complex patterns in construction and demolition (C&D) waste generation, considering factors like project type, location, and economic conditions [51]. LSTM can handle nonlinear relationships and multiple input features, improving prediction accuracy with large datasets. However, it requires more computational resources and tuning, and its complexity can make it less interpretable than simpler models [52].

Transformer-based models, such as Temporal Fusion Transformers (TFTs) [53], outperform ARIMA and LSTM in C&D waste forecasting by capturing long-range dependencies, handling multivariate inputs, and offering interpretability. Unlike ARIMA, they model nonlinearity, and unlike LSTM, they require less tuning while improving the accuracy in dynamic waste generation patterns.

6.2. Regression Models (Linear, Nonlinear, Ensemble)

6.2.1. Linear Regression

Machine learning applications for predicting construction and demolition (C&D) waste generation increasingly use various regression models to flexibly model the relationships between input variables and waste output. Linear regression [31], a fundamental statistical method, models the relationship between waste generation and independent variables like project size and duration. It is useful for simple predictions when relationships are approximately linear, offering an ease of interpretation and quantifiable coefficients for each predictor. However, it is limited to linear relationships and is sensitive to outliers, which can distort results [18].

6.2.2. Nonlinear Regression

Nonlinear regression models provide greater flexibility in capturing complex relationships between input variables and waste output in construction and demolition (C&D) waste generation. Key techniques include polynomial regression, which adds polynomial terms to model curvature, and Generalized Additive Models (GAMs) [54], which combine linear predictors with smooth functions. These methods are particularly effective when waste generation patterns vary significantly with project characteristics. Although they can yield more accurate predictions for nonlinear relationships, they tend to be more complex, require larger datasets for training, and may be less straightforward to interpret than linear models. Consequently, machine learning applications for predicting C&D waste generation are increasingly utilizing these nonlinear regression approaches [55].

6.2.3. Ensemble Models

Ensemble models are increasingly utilized in machine learning applications for predicting construction and demolition (C&D) waste generation, as they combine multiple models to enhance prediction accuracy and robustness [49]. Key techniques include Random Forests [39], which aggregate predictions from multiple decision trees to reduce overfitting; Gradient Boosting Machines (GBMs) [41,56], which build models sequentially to correct errors from previous iterations; and XGBoost, an optimized version of Gradient Boosting that is effective for structured data [57,58]. These methods excel in handling diverse input types and capturing complex interactions, making them popular for C&D waste predictions. While they typically offer high accuracy and robustness against overfitting and can effectively model nonlinear relationships, they are more computationally intensive, require additional tuning, and are less interpretable than linear regression, which may pose challenges in regulatory contexts where transparency is crucial [59].

Regression models, ranging from linear to nonlinear and ensemble approaches, play a crucial role in predicting C&D waste generation. Linear regression offers simplicity and interpretability, while nonlinear methods and ensemble models provide greater flexibility and accuracy in capturing complex relationships. Ensemble models, particularly XGBoost, outperform linear and nonlinear regression by effectively handling complex, nonlinear relationships, reducing overfitting, and improving accuracy. Their robustness, adaptability to diverse data, and superior predictive performance make them ideal for C&D waste forecasting despite their higher computational requirements. Combining these techniques with domain-specific insights and additional data sources can further enhance prediction capabilities in managing C&D waste effectively.

6.3. Hybrid Approaches for Predicting C&D Waste Generation

6.3.1. Combining Statistical Models with Machine Learning

Hybrid approaches for predicting construction and demolition (C&D) waste generation often involve combining traditional statistical models with machine learning techniques [40]. For many researchers, one effective method is to use statistical preprocessing, where regression analysis identifies key predictors and trends before applying more complex machine learning algorithms. For instance, multiple regression can elucidate relationships between variables, and this information can then serve as features for advanced models like Random Forests or neural networks [29,47]. This combination leverages the interpretability of statistical models alongside the predictive power of machine learning, potentially enhancing model performance by providing well-defined input features.

Hybrid models, namely Grey the Wolf Optimizer (GWO) with an ANN and the Aquila Optimizer (AOA) with an ANN exhibit superior predictive accuracy while demanding fewer input parameters compared to standalone ANN models [43]. This research underscores the significant potential of hybrid models, integrating metaheuristic techniques with machine learning methods, for effective C&DW management”.

Cha et al. [33] developed a combination of models to forecast the amount of debris generated during demolition in South Korean redevelopment areas. Decision trees, linear regression, KNN, and Principal Component Analysis (PCA) were all used in this model. Prediction performance was best for the decision tree model (R² = 0.872) and worst for the KNN (Chebyshev distance) model (R² = 0.627) when Principal Component Analysis was not used. The hybrid PCA-KNN model outperformed the decision tree models and the non-hybrid KNN model in terms of prediction accuracy (R² = 0.897). The values obtained using the k-nearest neighbors, PCA-KNN, and observed values were 987.06 (kg/m²), 993.54 (kg/m²), and 991.80 (kg/m²), respectively.

However, a drawback is that the approach may inherit the limitations of the initial statistical models, particularly if they do not adequately capture the underlying complexities of the data.

6.3.2. Ensemble Learning with Traditional Methods

Hybrid ensemble models for predicting construction and demolition (C&D) waste generation combine traditional predictive models, such as linear regression, with advanced ensemble learning methods like Random Forests or Gradient Boosting [54] to improve predictive accuracy. For example, predictions from a linear regression model can serve as an input to a Random Forest, allowing the ensemble model to refine these baseline predictions. This approach enhances robustness by leveraging diverse modeling techniques, enabling the capture of both linear trends and complex nonlinear relationships. However, it also introduces increased complexity in modeling and may lead to overfitting if not managed appropriately [60].

6.3.3. Machine Learning Post-Processing of Traditional Models

Hybrid approaches for predicting construction and demolition (C&D) waste generation often involve machine learning post-processing of traditional models, specifically through residual analysis and correction. In this method, a traditional model, such as ARIMA, generates initial predictions, and machine learning techniques are subsequently applied to the residuals—essentially the errors of these predictions—to enhance accuracy [14,46,61]. For example, after forecasting waste generation with ARIMA, a machine learning model like LSTM [62] can be trained on the residuals to capture additional patterns and reduce overall error. This approach leverages the strengths of both traditional forecasting and machine learning, allowing for adaptive learning from the initial model’s errors. However, it introduces complexity in ensuring that the models are effectively integrated and work well together.

6.3.4. Feature Engineering Using Domain Knowledge

A key component of hybrid approaches for predicting construction and demolition (C&D) waste generation involves using domain knowledge for feature engineering to enhance the creation of relevant features [21]. Traditional methods can help identify relevant features based on expert insights, which are then utilized in machine learning models. For instance, historical C&D project data can be analyzed to develop meaningful features, such as waste generation rates per project type, that are incorporated into the machine learning model [63]. This approach leverages expert knowledge, improving both the relevance and interpretability of the model inputs. However, it may require substantial domain expertise and data preprocessing efforts to ensure effective feature extraction. Kang et al. [27] constructed a theoretical framework that enables the collection, maintenance, and analysis of extensive data via Smart Building Information Modeling (BIM) that makes use of cutting-edge technologies like the Internet of Things (IoT). This framework can respond to user actions like waste quantitative evaluation, demolition procedure preparation, optimal disposal route selection, and waste management strategy execution. The findings demonstrate that the suggested framework would facilitate the development of environmentally friendly methods of waste disposal by offering decision-making and technical assistance to building industry planners and engineers.

6.4. Statistical Comparison of Predictive Machine Learning Models

The studies on predictive modeling for waste estimation and management highlight the importance of R and R² values in evaluating model performance. For instance, Cha et al. [40] combined Categorical Principal Components Analysis (CATPCA) with support vector machine regression (SVMR), achieving R² = 0.594 and R = 0.770, a substantial improvement over the baseline model (R² = 0.007, R = 0.083), showing how CATPCA enhanced model predictions. Coskuner et al. [2] demonstrated exceptionally high R² values for various waste categories, such as 0.95 for domestic waste, 0.99 for commercial waste, and 0.91 for construction and demolition (C&D) waste, reflecting the robustness of their models in predicting waste generation. In another study, Cha et al. [41] used Random Forest (RF) to predict demolition waste generation, with R² values ranging from 0.554 to 0.800 and Pearson’s R ranging between 0.691 and 0.871, indicating stable prediction performance despite the use of small datasets. Aknabi et al. [15] applied deep learning models, achieving an impressive R² = 0.97 on average, with category-specific values as high as 0.99 for landfill waste, underscoring the effectiveness of deep learning in predicting waste across various categories. Lu et al. [18] showed that multiple machine learning models, such as Multiple Linear Regression (MLR) and artificial neural networks (ANNs), delivered R² values ranging from 0.756 to 0.977, showing variability in model performance. The higher end of this range (e.g., 0.977) indeed suggests robust predictive accuracy, but the lower end (0.756) may still indicate room for improvement. Rema et al. [38] employed a CNN-LSTM model, achieving R² = 0.98, which highlights the model’s high accuracy in predicting material recovery post-demolition. Finally, Lu et al. [39] proposed a machine learning regression model for pre-renovation waste auditing, achieving R² = 0.83, demonstrating its effectiveness in renovation waste estimation across different regions. These studies collectively emphasize that high R² values (often exceeding 0.9) indicate models with strong predictive reliability, especially for waste generation and recovery predictions, while R values provide insight into the strength of the correlation between predicted and actual outcomes. Models with R² values above 0.7 are generally considered effective, and in applications like waste management, such high values ensure that predictive models can reliably inform decision-making. Both R and R² are crucial for assessing the performance of predictive models, with R² evaluating how well the model fits the data and R indicating the strength of the relationship between predictions and observed values.

7. Comparative Analysis of Machine Learning, AI, Economic Modeling, and Smart Technologies in C&D Waste Management

The studies by Cha et al. [40], Coskuner et al. [2], Soultanidis and Voudrias [20], Lu et al. [39], and Kang et al. [27] offer diverse approaches to demolition and construction waste management, with each utilizing distinct technologies and methodologies. Cha et al. [40] stand out for their innovative use of Categorical Principal Components Analysis (CATPCA) to improve support vector machine regression models for predicting demolition waste generation, a novel contribution that has not been explored before. In contrast, Coskuner et al. [2] employ artificial neural networks (ANNs), specifically multi-layer perceptron (MLP) models, to predict municipal solid waste generation from various sources, demonstrating a cost-effective and reliable AI-driven approach to waste management. Convolutional neural networks (CNNs) [15,25] have demonstrated significant potential in waste sorting and characterization by leveraging image analysis techniques to achieve high accuracy. However, their effectiveness is dependent on the availability of extensive training data and substantial computational resources. Cha et al. [41], also conducted a comparative analysis of Random Forest (RF) and Gradient Boosting Machine (GBM) models, showing that both are effective in predicting demolition waste using small datasets, offering practical tools for waste management in demolition projects. Soultanidis and Voudrias [20] take a more economic approach, focusing on the recycling costs of demolition waste in Greek residential buildings, with findings that the mixing of materials significantly impacts costs, providing useful insights for waste pricing and policy development. Lu et al. [39] introduce a machine learning regression model for pre-renovation construction waste auditing, offering a predictive tool that is adaptable to various regions and projects, unlike the other studies which focus on waste generation after construction begins. Finally, Kang et al. [28] integrate Building Information Modeling (BIM) with smart technologies, emphasizing a technological approach to waste management that not only improves efficiency but also offers cost savings, marking a departure from the more traditional modeling techniques in the other studies. Together, these studies highlight the varied methods—ranging from machine learning and AI to economic analyses and technological integration—used to address the challenges of waste prediction and management in demolition and construction, each contributing unique insights into the field.

8. Limitations, Practical Implications, and Future Research Directions

The application of machine learning (ML) in estimating, classifying, and predicting construction and demolition (C&D) waste has significant implications for sustainable waste management. The integration of ML models with existing waste management frameworks, such as Building Information Modeling (BIM), the Internet of Things (IoT), and Geographic Information Systems (GISs), can enhance decision-making by providing real-time insights into waste generation and material recovery. Furthermore, automated waste classification through image recognition and sensor-based monitoring can optimize sorting processes, reducing landfill dependency and promoting circular economy practices.

However, challenges such as data heterogeneity, regulatory variations, and the need for high-quality datasets hinder the full-scale adoption of ML-based waste management solutions.

Incorporating diverse and multi-source data can enhance model adaptability by integrating information from various regions, construction types, and regulatory contexts. Expanding the dataset in this manner enables the model to generalize effectively across different scenarios. Additionally, techniques such as data augmentation and synthetic data generation can mitigate data scarcity, ensuring robust model performance even in cases of limited real-world data.

Domain adaptation and transfer learning further improve model applicability by allowing pre-trained models to be fine-tuned with localized data. This approach reduces the need for extensive retraining while enabling models to adapt efficiently to new environments. By leveraging previously learned features, transfer learning enhances predictive performance and accelerates model deployment in diverse settings.

Adaptive and explainable models contribute to improved interpretability and reliability. Techniques such as SHAP values and feature importance analysis help identify the influence of regional and regulatory variations on model predictions. This understanding facilitates dynamic model updates, ensuring that predictive outputs remain transparent and aligned with evolving industry standards.

Continuous model updating is essential for maintaining model relevance over time. Periodic retraining with updated datasets ensures that models remain accurate despite changing regulations and evolving waste management practices. This approach minimizes performance degradation and enhances long-term model effectiveness.

Finally, robust validation strategies are crucial for assessing model stability and generalizability. Employing cross-validation across diverse data distributions and conducting external validation on independent datasets help evaluate the model’s reliability in different contexts. These validation techniques strengthen confidence in the model’s predictive capabilities, ensuring its applicability in real-world waste sorting and characterization tasks.

To address these limitations, future research should focus on developing standardized data collection protocols, incorporating domain adaptation techniques to improve model generalizability across different regions, and leveraging federated learning approaches to enable privacy-preserving collaborations across industries. Additionally, integrating ML models with life-cycle assessment (LCA) frameworks can provide a more holistic understanding of the environmental impact of C&D waste. Research into hybrid AI models, combining deep learning with optimization algorithms, can further improve waste prediction accuracy while ensuring computational efficiency.

By advancing these areas, ML-driven methodologies can play a pivotal role in transforming C&D waste management, fostering sustainable construction practices, and supporting data-driven policy interventions for a more resource-efficient future.

9. Conclusions

• Introduction to C&D Waste Management Challenges: Effective management of construction and demolition (C&D) waste poses significant environmental, economic, and operational challenges. Traditional methods, including manual surveys, visual inspections, statistical analyses, and Building Information Modeling (BIM), provide insights but lack the precision and scalability needed for modern systems.
• Emergence of Machine Learning in Waste Management: Machine learning (ML) techniques have revolutionized waste management, offering advanced predictive modeling capabilities. Models like Random Forest (RF), Gradient Boosting Machines (GBMs), artificial neural networks (ANNs), and convolutional neural networks (CNNs) consistently outperform traditional methods in accuracy and reliability.
• Predictive Accuracy of ML Models: ML models achieve R² values often exceeding 0.9, demonstrating robust predictive capabilities that are essential for strategic decision-making in waste management. Hybrid approaches, such as combining Categorical Principal Component Analysis (CATPCA) with support vector machine regression (SVMR), further enhance predictive performance, particularly in demolition waste generation and material recovery.
• Integration of Large and Unstructured Data: Advanced ML frameworks excel in handling large datasets and real-time inputs, including unstructured data like images, which is crucial for material classification tasks. These frameworks optimize waste estimation, recovery, and cost-effective data-driven strategies that are aligned with sustainability goals.
• Challenges in ML-Driven Waste Management: Despite the advantages, challenges such as data heterogeneity, computational demands, and integration barriers remain. Overcoming these obstacles requires standardized data collection protocols, enhanced computational tools, and cross-sector collaboration among stakeholders.
• The Role of ML in Promoting Sustainable Practices: This analysis underscores the transformative potential of ML in C&D waste management. By providing precise, scalable, and adaptable solutions, ML-driven methodologies support the transition to a circular economy, promote sustainable construction practices, and aid evidence-based policy development.

This comparative analysis underscores the transformative potential of ML in addressing the challenges of C&D waste management. By delivering precise, scalable, and adaptable solutions, ML-driven methodologies facilitate the transition to a circular economy, promote sustainable construction practices, and support evidence-based policy development. The integration of ML technologies into waste management systems represents a paradigm shift, enabling a more sustainable, efficient, and resilient approach to managing the complexities of C&D waste.