Self-Supervised Learning Framework for Automated Material Grouping of Demolition Waste: Advancing Sustainable Waste Management
Open accessAbstract
The construction and demolition (C&D) sector generates significant waste, necessitating efficient sorting and recycling to promote sustainable waste management. Traditional sorting methods in C&D waste management often depend on manual processes, which are labor-intensive and prone to errors, thus limiting recycling efficiency. While automated sorting systems have been introduced, they face challenges with the complex, heterogeneous nature of C&D debris and require large datasets that are manually labeled—meaning every image must be tagged with material types by experts—for training. This study addresses this gap by developing a self-supervised learning framework that reduces reliance on labeled data for effective feature extraction and material grouping. We implemented contrastive learning and autoencoder models, enhancing model performance through a dual approach of fine-tuning and parameter optimization, including edge detection, temperature, and batch size adjustments. The contrastive learning model, when optimized with lower temperatures and smaller batch sizes, exhibited superior feature differentiation and minimized loss. Clustering results highlighted that agglomerative hierarchical clustering provided the most coherent material groupings, outperforming other methods in adaptability to diverse debris characteristics. This framework provides a scalable, autonomous method for material grouping, which can streamline the sorting process by categorizing materials based on shared characteristics. By facilitating the initial grouping of materials, it reduces the manual effort required for detailed sorting and enhances recycling efficiency. Our findings demonstrate the effectiveness of self-supervised learning models in identifying distinct material features and clustering similar materials, contributing to sustainable waste management by enabling more efficient material recovery and recycling processes.
AI snapshot
Written from the paper’s full text.
TL;DR
This paper develops a self-supervised learning framework using contrastive learning and autoencoders to group demolition waste materials without manual labels. It finds that contrastive learning with strong edge detection, lower temperature, and smaller batch sizes produces the best feature separation, and that agglomerative hierarchical clustering yields the most coherent material groupings.
Key findings
- Contrastive learning with fine-tuning and strong edge detection achieved the largest loss reduction, stabilizing around 0.1; without fine-tuning, strong and soft edge detection converged at higher loss levels, with soft edge detection performing worst.
- Among tested temperatures from 0.1 to 0.5, a temperature of 0.1 achieved the lowest final contrastive loss, nearing 0.04 by the end of training.
- Smaller batch sizes of 8 and 16 achieved lower loss values more rapidly than larger sizes; batch sizes larger than 32 led to slower convergence and higher final losses.
- Agglomerative hierarchical clustering outperformed K-means, Gaussian mixture models, and spectral clustering, with spectral clustering showing the weakest separation; contrastive learning also produced more coherent AHC clusters than the autoencoder.
Methods
- A dataset of 300 demolition debris images was collected in a controlled, small-scale laboratory setup at 1:14 scale, featuring metals, small-size concrete, and large-size concrete pieces.
- Preprocessing included Python-based thresholding for background removal, resizing to 1440 x 1084 pixels, and two edge/texture treatments: high-contrast enhancement with strong “FIND_EDGES” filtering and contrast enhancement with softer “CONTOUR” filtering.
- The framework used a MobileNetV2-based contrastive learning model with a projector head and an autoencoder with MSE reconstruction loss; learned features were clustered using K-means, agglomerative hierarchical clustering, Gaussian mixture models, and spectral clustering, with evaluation by loss functions and expert visual inspection.
Limitations
- The paper states that future work should test scalability in real-world applications, larger datasets, and diverse environments, and should optimize clustering for complex, overlapping, or noisy data distributions.
- The paper also notes that agglomerative hierarchical clustering faces scalability challenges with larger datasets and that spectral clustering is computationally demanding with high-dimensional data.
- Evident from the text: the study uses only 300 images from a controlled 1:14 laboratory setup, and clustering performance is assessed through expert visual inspection rather than quantitative clustering metrics or labeled ground-truth material categories.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
1 of 5 cited sources checked · 0 supported · 0 flagged · accuracy not yet measured
- Findingown data
The model configuration with fine-tuning and strong edge detection achieved the most significant reduction in contrastive loss, stabilizing around 0.1.
“The model configuration with fine-tuning and strong edge detection achieved the most significant reduction in loss, stabilizing around 0.1, indicating that this setup is particularly effective for capturing detailed features in the debris images.”
Quote in the paperSubstantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingown data
The autoencoder model displayed slightly less distinct clusters than the contrastive learning model.
“In comparison, the autoencoder model (Figure 5b) displayed slightly less distinct clusters.”
Quote in the paperNot a substantial claim- Rests on an assumption: Visual inspection by human experts is a valid measure of clustering coherence and quality.
- Findingown data
Lower temperatures contribute to better feature differentiation in contrastive learning.
“Further analysis of the effect of temperature on contrastive loss (Figure 2b) shows that lower temperatures contribute to better feature differentiation.”
Quote in the paperNot a substantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingown data
Self-supervised learning models are effective in identifying distinct material features and clustering similar materials.
“Our findings demonstrate the effectiveness of self-supervised learning models in identifying distinct material features and clustering similar materials, contributing to sustainable waste management by enabling more efficient material recovery and recycling processes.”
Quote in the paperSubstantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingown data
Gaussian mixture models struggled with overlapping clusters due to mixed material textures and inconsistencies, resulting in less distinct groupings.
“Gaussian mixture models (Figure 3b), while providing flexibility through probabilistic modeling, struggled with overlapping clusters due to mixed material textures and inconsistencies, resulting in less distinct groupings.”
Quote in the paperNot a substantial claim- Rests on an assumption: Visual inspection by human experts is a valid measure of clustering coherence and quality.
- Findingown data
Agglomerative hierarchical clustering provided the most coherent material groupings, outperforming other methods in adaptability to diverse debris characteristics.
“Clustering results highlighted that agglomerative hierarchical clustering provided the most coherent material groupings, outperforming other methods in adaptability to diverse debris characteristics.”
- Also stated in: § 4 Findings and Discussion · plain
“Agglomerative hierarchical clustering (Figure 3d) outperformed the other methods by consistently grouping similar materials with greater coherence.”
- Also stated in: § 5 Conclusions and Further Research · plain
“The evaluation of clustering algorithms highlighted agglomerative hierarchical clustering as the most effective method for achieving coherent material groupings, offering a robust approach for handling heterogeneous waste materials.”
Quote in the paperSubstantial claim- Rests on an assumption: Visual inspection by human experts is a valid measure of clustering coherence and quality.
- Also stated in: § 4 Findings and Discussion · plain
- Findingown data
Spectral clustering showed the weakest performance, with poor separation among clusters.
“Spectral clustering (Figure 3c) showed the weakest performance, with poor separation among clusters.”
Quote in the paperNot a substantial claim- Rests on an assumption: Visual inspection by human experts is a valid measure of clustering coherence and quality.
- Findingown data
Self-supervised learning models, particularly contrastive learning and autoencoders, are effective in capturing meaningful features for material grouping in construction and demolition debris.
“This research demonstrated the effectiveness of self-supervised learning models, particularly contrastive learning and autoencoders, in capturing meaningful features for material grouping in construction and demolition debris.”
Quote in the paperSubstantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingown data
Higher temperatures maintained higher final loss values, indicating less precise clustering and separation of material types.
“Higher temperatures, while still reducing loss, maintained higher final values, indicating less precise clustering and separation of material types.”
Quote in the paperNot a substantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingown data
The contrastive learning model, especially when fine-tuned with strong edge detection and optimized hyperparameters such as lower temperatures and smaller batch sizes, outperformed the autoencoder in loss minimization and feature differentiation.
“The contrastive learning model, especially when fine-tuned with strong edge detection and optimized hyperparameters such as lower temperatures and smaller batch sizes, outperformed the autoencoder in terms of loss minimization and feature differentiation.”
Quote in the paperSubstantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingreasoning
Smaller batch sizes of 8 and 16 achieved lower loss values more rapidly compared to larger batch sizes.
“Smaller batch sizes of 8 and 16 achieved lower loss values more rapidly compared to larger sizes.”
Quote in the paperNot a substantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Findingown data
Among tested temperature values (0.1 to 0.5), a temperature of 0.1 achieved the lowest final loss, nearing 0.04 by the end of training.
“Among the tested values (0.1 to 0.5), a temperature of 0.1 achieved the lowest final loss, nearing 0.04 by the end of training.”
Quote in the paperNot a substantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
- Interpretationreasoning
This approach demonstrates the value of self-supervised learning in scenarios with limited labeled data.
“This approach contributes to the fields of machine learning and sustainable waste management, demonstrating the value of self-supervised learning in scenarios with limited labeled data.”
Quote in the paperSubstantial claim - Interpretationreasoning
The autoencoder's reconstruction-based approach may be less responsive to the nuanced variations necessary for accurate debris categorization.
“The autoencoder, while effective, converged more gradually, suggesting that its reconstruction-based approach may be less responsive to the nuanced variations necessary for accurate debris categorization.”
Quote in the paperSubstantial claim - Interpretationreasoning
Agglomerative hierarchical clustering effectively captured both fine-grained and large-scale similarities, making it well-suited for the complex and heterogeneous nature of C&D waste.
“This bottom-up approach, which iteratively merges clusters based on proximity, effectively captured both fine-grained and large-scale similarities, making it particularly well-suited for the complex and heterogeneous nature of C&D waste.”
Quote in the paperSubstantial claim - Interpretationreasoning
The success of the framework supports the development of an autonomous system capable of enhancing material sorting accuracy without significant effort for data labeling.
“This success supports the development of an autonomous system capable of enhancing material sorting accuracy without significant effort for data labeling, thereby contributing to automated, sustainable waste management operations.”
Quote in the paperSubstantial claim- Rests on an assumption: The clustering algorithms' performance on the 300-image dataset generalizes to real-world C&D waste sorting.
- Interpretationreasoning
The scarcity of high-quality labeled data poses a significant barrier to scaling supervised learning models for real-world waste sorting applications.
“Consequently, the scarcity of high-quality labeled data poses a significant barrier to scaling supervised learning models for real-world waste sorting applications.”
Quote in the paperSubstantial claim - Premisecitation
Developing labeled datasets for C&D debris is challenging due to the complex and heterogeneous nature of waste materials.
“Developing such datasets for C&D debris is challenging due to the complex and heterogeneous nature of waste materials (Ji et al., 2023).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Ji et al., 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- May combine two claims. 11
- The quote only partly states this claim. 1
- Classified as probably not a substantial claim. 3
May combine two claims.
PremisecitationAutomated sorting systems enhance sorting accuracy, improve recovery rates, and reduce operational costs.
“Automated systems offer a promising alternative by enhancing sorting accuracy, improving recovery rates, and reducing operational costs, thus aligning with broader environmental sustainability goals (Llatas, 2011).”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Llatas, 2011 source
Section 5.6 (p. 30) and the Conclusions (p. 35) say that detailed, project-specific waste classification could improve recovery. They do not address automated sorting systems or their effects on sorting accuracy, recovery rates, or operational costs; searches for automated sorting and cost savings found no such passage. (LLM, conf=0.76)
Experimental — accuracy not yet measured
May combine two claims.
Findingown dataThe contrastive learning model converged faster and achieved a lower final loss than the autoencoder.
“Comparing the performance of the contrastive learning model with the autoencoder (Figure 2d) revealed that the contrastive learning model not only converged faster but also achieved a lower final loss.”
Quote in the paperSubstantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
May combine two claims.
FindingreasoningBatch sizes larger than 32 led to slower convergence and higher final losses.
“In contrast, batch sizes larger than 32 led to slower convergence and higher final losses, possibly due to reduced variability within each batch, which limits the model's ability to capture the full diversity of features in the debris images.”
Quote in the paperNot a substantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
The quote only partly states this claim.
FindingreasoningThe contrastive learning model exhibited superior clustering coherence, with distinct and well-separated groups, compared to the autoencoder model.
“The contrastive learning model (Figure 5a) exhibited superior clustering coherence, with distinct and well-separated groups.”
Quote in the paperSubstantial claim- Rests on an assumption: Visual inspection by human experts is a valid measure of clustering coherence and quality.
May combine two claims.
Findingown dataK-means clustering offered clear cluster boundaries but was sensitive to the specified number of clusters and initial centroid selection, reducing adaptability when handling complex and variable materials.
“K-means clustering (Figure 3a) offered clear cluster boundaries but was sensitive to the specified number of clusters and initial centroid selection, reducing adaptability when handling complex and variable materials.”
Quote in the paperNot a substantial claim- Rests on an assumption: Visual inspection by human experts is a valid measure of clustering coherence and quality.
May combine two claims.
Findingnone givenThe contrastive learning model, when optimized with lower temperatures and smaller batch sizes, exhibited superior feature differentiation and minimized loss.
“The contrastive learning model, when optimized with lower temperatures and smaller batch sizes, exhibited superior feature differentiation and minimized loss.”
Quote in the paperSubstantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
May combine two claims.
Findingown dataWithout fine-tuning, both strong and soft edge detection approaches converged at higher loss levels, with soft edge detection performing the worst.
“Without fine-tuning, both strong and soft edge detection approaches converged at higher loss levels, with soft edge detection performing the worst.”
Quote in the paperNot a substantial claim- Rests on an assumption: Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
May combine two claims.
PremisecitationManual labeling of C&D waste materials is labor-intensive, costly, and prone to inconsistencies, exacerbated by frequent variations in debris composition and state.
“Manual labeling of these materials is labor-intensive, costly, and prone to inconsistencies, exacerbated by frequent variations in debris composition and state (Hassan et al., 2022).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Hassan et al., 2022 source
May combine two claims.
InterpretationreasoningBy facilitating the initial grouping of materials, the framework reduces the manual effort required for detailed sorting and enhances recycling efficiency.
“By facilitating the initial grouping of materials, it reduces the manual effort required for detailed sorting and enhances recycling efficiency.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationown dataThe autoencoder's latent representations were less discriminative than those derived from contrastive learning, resulting in less coherent clustering.
“While the autoencoder captured underlying data features, its latent representations were less discriminative than those derived from contrastive learning, resulting in less coherent clustering.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim. May combine two claims.
Interpretationnone givenThe automatically grouped data produced by this framework can be utilized as labeled datasets to train subsequent machine learning models, enabling automated waste sorting operations for material recovery.
“The results of this framework, which yield automatically grouped data, can be utilized as labeled datasets to train subsequent machine learning models, enabling automated waste sorting operations for material recovery.”
Quote in the paperSubstantial claim- Rests on an assumption: The clustering algorithms' performance on the 300-image dataset generalizes to real-world C&D waste sorting.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationThe construction and demolition sector accounts for over one-third of all solid waste generated worldwide.
“The construction and demolition (C&D) sector significantly contributes to global waste production, accounting for over one-third of all solid waste generated worldwide (Eurostat, 2020).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Eurostat, 2020
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
May combine two claims.
PremisecitationConventional manual sorting methods in C&D waste management are labor-intensive, inefficient, and prone to human error, resulting in material contamination and decreased recycling efficiency.
“Traditional sorting methods in C&D waste management often depend on manual processes, which are labor-intensive and prone to errors, thus limiting recycling efficiency.”
- Also stated in: § 1 Introduction · plain
“However, conventional manual sorting methods are laborintensive, inefficient, and prone to human error, resulting in material contamination and decreased recycling efficiency (Coelho and de Brito, 2013).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Coelho and de Brito, 2013
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Also stated in: § 1 Introduction · plain
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenThe sample size of 300 images was selected to ensure sufficient diversity in material types while maintaining a realistic level of manual effort for preprocessing.
“The sample size of 300 images was selected to ensure sufficient diversity in material types while maintaining a realistic level of manual effort for preprocessing.”
Quote in the paperNot a substantial claim
The paper relies on this without support.
The 300-image dataset collected in a 1:14 scale laboratory setup is representative of typical C&D waste.
“A dataset of 300 images of demolition debris, representative of typical C&D waste, was collected in a controlled, smallscale experimental setup (i.e., 1:14 scale) conducted as a laboratory experiment.”
The paper relies on this without support.
Visual inspection by human experts is a valid measure of clustering coherence and quality.
“Clustering performance was assessed through visual inspection, where human experts reviewed the coherence and quality of the material groupings.”
The paper relies on this without support.
Minimization of contrastive loss and reconstruction loss reflects meaningful feature learning for material grouping.
“For the autoencoder, reconstruction loss (mean squared error) was used to measure how accurately the model could reconstruct input images, with lower loss values indicating that the model had learned key features needed for effective material representation.”
“Similarly, contrastive loss was used for the contrastive learning model, with the minimization of this loss reflecting the model's ability to distinguish between similar and dissimilar materials, leading to meaningful feature learning.”
The paper relies on this without support.
The material types and configurations in the dataset (metals, small-size concrete, large-size concrete) are representative of the variability in real-world C&D debris.
“The images featured various C&D waste materials, including metals, small-size concrete, and large-size concrete pieces, arranged in diverse configurations to capture the variability inherent in demolition debris (Figure 1a).”
The paper relies on this without support.
The clustering algorithms' performance on the 300-image dataset generalizes to real-world C&D waste sorting.
“By iteratively refining the models and clustering process based on both quantitative (loss functions) and qualitative (visual inspection) feedback, we ensured that the final framework was optimized for practical, real-world application in waste management.”
Cites
The publisher hasn’t deposited this paper’s reference list.
- Authors
- Alshami, A., & Choi, J.
- Affiliations
- Florida A&M University - Florida State University College of Engineering (US)
- Venue
- CIB Conferences
- Publisher
- Purdue University (bepress)
- Published
- 2025-06-19
- Volume
- 1
- Issue
- 1
- ISSN
- 3067-4883 (ISSN-L)
- OpenAlex
- W4411472635
- Cited by
- 0
- References
- 0
- FWCI
- 0.00
- Type
- conference paper
- Language
- en
- Open access
- green
- docs.lib.purdue.edu
- Locations
- CIB Conferences (publishedVersion)
- Purdue e-Pubs (Purdue University System) (submittedVersion)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Engineering › Building and Construction › Recycled Aggregate Concrete Performance
- Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
- Keywords
- construction and demolition waste, automated waste sorting, self-supervised learning, contrastive learning, autoencoder, agglomerative clustering, waste recycling, sustainable waste management
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-16.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-16.
Cited sources are checked by gpt-6-sol.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
