Application of Machine Learning in Plastic Waste Detection and Classification: A Systematic Review

Abstract

The intersection of artificial intelligence and environmental sustainability has become a relevant exploration domain in the contemporary era of rapid technological advancements and complex global challenges. This work reviews the application of machine learning (ML) models to address the pressing issue of plastic waste (PW) management. By systematically examining the state of the art with snowballing, this research aims to determine the efficiency and effectiveness of ML-based methods for PW detection and classification. Considering the increasing environmental concerns and information processing potential, this article hypothesised that ML models could contribute to more sustainable PW management practices. For this purpose, two scientific article repositories were examined from 2000 to 2023, and 188 articles were identified. After the systematic screening procedure, 28 were selected. Additionally, 28 more articles were included by snowballing. It was observed that accuracy in either detection or classification problems often exceeded the 80% detection accuracy benchmark, further improving when the model combination was employed. As a result, strong support was reached for the applicable potential of ML in PW. It was also concluded that models based on convolutional neural networks were the most commonly used.

1. Introduction

The urgent need for environmental sustainability has brought about remarkable global efforts to promote recycling on a larger scale. Governments, organisations, and individuals worldwide have acknowledged the importance of recycling to reduce plastic waste (PW) and conserve resources.

Advancements in technology, infrastructure, and policy frameworks have accompanied the evolution of recycling practices. Today, recycling programs encompass a wide range of materials, and recycling has become a symbol of hope for our planet. Artificial intelligence (AI) will be vital in addressing challenges and increasing recycling rates worldwide [1,2].

To explore comprehensively the challenges and opportunities presented by integrating machine learning (ML) approaches into PW reduction strategies, we must first characterise the context of the problem at hand. The global PW problem is a considerable challenge today, casting a shadow over environmental sustainability, economic efficiency, and social well-being. This multi-faceted issue transcends geographical boundaries, affecting communities, industries, and ecosystems worldwide [3].

It was estimated that currently, only 9% of global PW finds its way to effective recycling, 12% is incinerated, and 79% is accumulated in landfills or in the natural environment. PW, in particular, accounts for a staggering 91% that has yet to be reclaimed [3,4,5]. These statistics portend a dire future, with projections indicating that by 2050, oceans may be burdened with more plastic than fish by weight [6].

The gravity of this environmental crisis necessitates a concerted and innovative response that uses AI and ML’s transformative potential. These technologies offer promising possibilities for PW reduction and are essential to revolutionising how we approach PW management, recycling, and resource allocation. It is therefore necessary to explore the complex issue of global trash and investigate ML-driven solutions to address the problem [7].

The problem of global waste is closely related to the widespread use of plastics in our daily lives. Although plastics have brought unprecedented convenience, durability, and versatility to various industries and applications, they have also led to significant environmental problems.

One of the biggest challenges in waste management (WM) is the increasing use of non-recyclable plastics. These materials are commonly used for single-use items or packaging and are difficult to recycle using traditional methods. Their persistence in the environment worsens pollution and makes it harder to reduce waste in landfills or through incineration [8].

Efficient WM depends on a strong infrastructure that covers the processes of collection, sorting, recycling, and disposal. However, many areas worldwide lack the necessary infrastructure to manage the increasing amount of PW. This inadequacy leads to the leakage of plastics into natural ecosystems and oceans, causing long-term environmental damage [9].

Plastics can be classified into seven primary categories based on their formulation. Each category has a unique Resin Identification Code (RIC) and specific properties affecting recyclability and disposal methods. For instance, Polyethylene Terephthalate (PETE) is often used in beverage bottles and food containers and is easy to recycle. On the other hand, High-Density Polyethylene (HDPE) is known for its strength and is used in containers and piping. Polyvinyl Chloride (PVC) is commonly used in construction materials, Low-Density Polyethylene (LDPE) is found in plastic bags, Polypropylene (PP) is used in automotive parts and textiles, and Polystyrene (PS) is used in insulation and packaging. Finally, the “Other” category includes various plastics such as polycarbonate and bioplastics. Each category presents distinct recycling and disposal challenges, requiring nuanced understanding and approaches to manage and mitigate their environmental impact. This diversity emphasises the complexity of waste management efforts and highlights the importance of tailored strategies to address each type of plastic’s specific needs and challenges [8].

The problem of how to recycle these was addressed in multiple initiatives worldwide. Specifically, Operation Green Fences, initiated by Chinese authorities, highlighted the need for stricter quality controls in the recycling industry. The operation exposed the consequences of lax standards in the global trade of recyclables, illustrating the importance of responsible recycling practices [10,11]. As expected, solving the PW problem requires innovative technology. AI has emerged with promising tools for reducing waste and increasing recycling. These technologies can potentially revolutionise PW sorting, classification, and resource allocation processes, leading to a cleaner and more sustainable planet.

The intersection of AI and environmental sustainability has become a relevant exploration domain in the contemporary era of rapid technological advancements and complex global challenges. In recent years, substantial progress has been made in applying ML models to the pressing issue of PW management, with studies showing promising results in waste detection and classification accuracies. Especially, we can observe a transition from conventional machine learning models, such as support vector machines, to newer and more complex models based on deep learning models. These advancements suggest that ML technologies are approaching the threshold of practical applicability in real-world waste management contexts, potentially revolutionising how we approach PW management, recycling, and resource allocation. Although previous reviews were performed on general PW, to the best of the authors’ knowledge, a systematic review specific to PW analysis with detection and classification methodologies was not previously performed.

The main objective of this work is to explore the abilities of ML models in developing efficient strategies for managing PW, which is a growing environmental concern. This article is based on the concept that ML models can transform the processes of identifying, sorting, and recycling PW. To examine this proposal, two research questions were formulated: Can ML models achieve PW detection and classification accuracy suitable for real-world applications? Which ML approach is more suitable for PW detection?

Guided by this central aim, the article is structured around the following goals: To identify and examine the leading ML models through a systematic review, highlighting their capabilities in image-based PW detection and classification. To gather and analyse existing data to examine the previously identified ML models in PW tasks, establishing a performance benchmark.

By pursuing these goals, this work seeks to provide a further understanding of the role of ML in environmental sustainability. This article is structured as follows: Section 2 outlines the search strategy used to identify the initial articles for evaluation. It also covers the eligibility criteria during the screening stage and concludes with the analysis method and the articles included. Section 3 describes the search results and thoroughly analyses the included articles. Section 4 highlights the conclusions of the article by describing the major findings, while Section 5 outlines the main challenges and future research directions.

2. Materials and Methods

This section presents the methods used for retrieving and analysing articles. We have performed the analysis based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to allow the reproducibility of our results [12].

2.1. Search Strategy

This work searched articles across two databases: Web of Science Core Collection (WSCC) and Institute of Electrical and Electronics Engineers Xplore (IEEEX). WSCC provides a comprehensive and multidisciplinary database that offers access to indexed journals across various fields. The IEEE Xplore Digital Library is a specialised database focusing on electrical engineering, computer science, and electronics. Cumulatively, these databases ensured a thorough search due to their extensive coverage of multiple fields and publishers, allowing for a comprehensive examination of the topic under analysis.

The article search was carried out on 17 December 2023 and was filtered to scan only the article’s title, abstract, and author-defined keywords published in the 2000–2023 timeframe. The search string “artificial intelligence AND machine learning AND recycling” was utilised to filter and narrow the search results according to the topic of interest.

The keywords were selected to maximise the retrieval of pertinent information while minimising the loss caused by adjectival usage. Specifically, the keyword “recycling” was used to ensure that all search results were related to recycling. In addition, and to ensure the inclusion of standard ML models, the keywords “machine learning” and “artificial intelligence” were used together with the “AND” operator.

2.2. Systematic Search

The systematic article selection process is depicted in the PRISMA diagram shown in Figure 1. A total of 188 articles were found by searching two different databases, and the specifics are presented in Figure 1. The WSCC had the highest number of publications, with 120 articles, while 68 articles were found in IEEEX.

A duplicate record elimination process was conducted before passing the articles to the initial screening phase. Nine articles were removed due to duplication. Then, three independent scorers evaluated the relevance of each article. The inclusion criteria were “articles that included ML or general AI applied to WM”. The exclusion criteria were “articles not focused on WM that did not have either ML or AI mentioned” and “articles not written in English”.

During this procedure, a voting system was employed. Each scorer evaluated the title and abstract of each article and voted for inclusion, exclusion, or further discussion. Articles receiving two votes for inclusion were automatically included, while those receiving two votes for exclusion were automatically excluded. Of the 179 screened articles, 143 were excluded, thus resulting in 36 articles in the second screening.

Two articles were excluded during the second screening analysis, which involved the complete article screening process, due to their lack of specificity of waste type unrelated to PW [13,14]. Another two were excluded because despite having a broad Focus on WM, they had a limited specificity to plastic WM [15,16]. Finally, four were excluded because the reports were scattered on diverse topics related to recycling and WM [17,18,19,20]. The selection process resulted in a total of 28 articles that were included in the systematic review [5,15,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46].

The resulting 28 studies included in this review all considered AI methods to address plastic WM, focusing on reduction and recycling. However, there is a considerable difference among the articles regarding their use of ML models in analysis. Some articles focused on practical and objective analysis, while others adopted a more holistic approach to evaluate the current state of the art. To simplify the analysis, we decided to group them based on the ML models they used and their accuracy in detecting and/or classifying waste.

3. Results and Discussion

Based on an analysis of the publication year of the 28 included articles, it was observed that research activity on the studied subject began five years ago. This is shown in Figure 2, which displays the distribution of published articles by year. The systematically reviewed articles started in 2019 (3.57%), and the gradual increase in the number of publications after that year indicates a growing interest in the topic. Five articles (17.86%) were published in 2020, five articles (17.86%) were published in 2021, ten articles (37.71%) were published in 2022, and seven articles (25.00%) were published in 2023. The peak in 2022 suggests that the subject has recently attracted significant attention. The prevalence of this trend emphasises the contemporary significance of the examined subject matter, stressing the need for this review to consolidate knowledge and point out new directions for future research.

We then checked the cited literature in the articles, snowballing the additional literature, especially from previous non-systematic reviews [5,23]. These reviews pointed out the relevance of deep learning-based models for PW detection and classification. The additional articles were only included if they were published with peer review and if they explicitly mentioned the usage of the method for plastic recycling. This way, it was possible to include 28 additional articles [47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74]. These results are presented in Figure 2, showing that the total number of examined articles was 56.

The following subsections analyse the results summarised in

Table 1. Analysis of the reviewed articles that have reported an accuracy-based metric.

Study	Year of Publication	Number of Categories/Classes	Detection Accuracy (Specific Method—Accuracy in %)	Classification Accuracy (Specific Method—Accuracy in %)
[25]	2020	4	YOLOv3—85.00 +	-
[26]	2020	3	MobileNetV2—86.23 #	-
[39]	2022	2	SSD-MobileNet-V1—63.64 #	-
[44]	2022	5	EfficientDet—67.40	-
[30]	2023	6	Resnet50—87.00 [53] $ Resnet50—88.00 [52] $ VGG19—88.00 [54] $ ConvoWaste—98.30	-
[24]	2020	-	-	ResNet-34—89.96
[45]	2020	2	-	CNN—89.00 [48] $ CNN—98.20 [47] $ AlexNet and GoogleNet and ResNet-50—99.95
[22]	2021	6	-	AlexNet—75.00 InceptionV1—82.00
[15]	2022	4	-	Deep CNN4—37.00 Deep CNN5—56.70
[43] *	2022	4	-	SqueezeNet—66.84 AlexNet—68.13 InceptionNet—74.41 ResNet—76.59 MobileNet_V2—76.77 GoogleNet—76.89 VggNet—77.78 EfficientNet—79.49 DenseNet—80.63
[32]	2022	4	-	CNN—95.63
[21]	2022	4	-	CNN—80.88 [50] $ VGG16—88.42 [49] $ MLB-DCNN—92.60 [51] $ FNN-TH—97.02
[46] ~	2022	5	-	CNN—95.60 YOLOv5—98.30
[38]	2022	3	-	Resnet-50—96.50 InceptionV3—98.60 MobileNetV2—99.60
[42]	2023	4	-	InceptionV3—36.00 ResNeX50—45.20 VGG-16—46.50 ResNet50—47.85 ResNet—52.44
[36]	2023	8	-	EfficientNet-B3—97.32 CNN—98.50 [55] $
[5] †	2021	-	CNN—64.00 [71] $ R-CNN—74.10 [69] $	AlexNet—83.00 [74] $ VGG16—93.00 [73] $ CNN—93.50 [70] $ Capsule-Net—93.60 [72] $ Capsule-Net—95.80 [72] $
[23] †	2022	-	Tiny-YOLO—31.60 [68] $ YOLOv2—47.90 [68] $ SSD—67.40 [68] $ YOLO-Green—78.04 [65] $ Faster RCNN—81.00 [68] $ ResNet-50—81.48 [66] $ L-SSD—83.48 [67] $	YOLOv5—73.20 [59] $ CNN—92.20 [63] $ EfficientDet—92.87 [58] $ InceptionV3—93.13 [60] $ AlphaTrash—94.00 [64] $ ThanosNet—94.70 [61] $ GCNet—97.54 [56] $ DNN-TC—98.00 [62] $ DSCAM—98.90 [57] $

* Only the results of the models with pre-training are presented. ^~ Only the results of the model with the 80–20% approach are presented, and we are only reporting the CNN-based models. ^# Reported as Mean Average Precision (mAP). ⁺ Reported as recognition rate. ^$ Results of an article identified by snowballing. ^† Review article.

, which provides an overview of these articles, including how many classes/categories were considered. The accuracy is presented in a broader way as an initial examination combining different methodologies used in different articles. It was observed that most works use a standard convolutional neural network (CNN) architecture or a model that is based on it, such as Single-Shot MultiBox Detector (SSD) and You Only Look Once (YOLO). Therefore, two groups were formed, specifically, the articles that used these ML models and those that used other approaches. When the used architecture was named, it was opted to specify it; otherwise, it is just indicated as CNN.

It is important to note that the additional literature included by snowballing was used only to clarify the performance of the considered models and is not intended to be included in the in-depth subsequent analysis since the articles were not identified systematically. Additionally, the articles identified by snowballing (marked with $ and the reference in front) are included in the same row of the article that was identified by the systematic search to facilitate the replicability of the search. No duplicated snowballing-identified articles were included, and the number of classes/categories and year of publication indicated in the row are unrelated to these articles.

Furthermore, three review articles were included but did not specifically examine the considered models. Specifically, Seyyedi et al. [37] studied AI-based systems concerning marine plastics and circular economy. There is also a review by Shennib and Schmitt [31] regarding circular economy and WM systems using AI and a review on enzyme-embedded biodegradable agricultural plastics by Maraveas et al. [35]. Several parameters were extracted from the various articles, and the data from the different articles are presented in Table 1.

A bibliometric analysis was also performed. Of the reviewed 56 articles, 27 were published in international conferences (13 in the systematically reviewed articles) and 29 in journals (15 in the systematically reviewed articles). However, conference articles that were published in a journal paper (for example, the Journal of Physics: Conference Series) were counted as journal articles. The analysis ended up as a balanced mix between international conferences and journal publications. Regarding the number of citations indicated in Google Scholar, for the systematically reviewed articles (on 16 June 2024), it was 1183 for the systematically reviewed articles and 1693 for the snowballed articles, leading to a total of 2876 citations, highlighting the relevance of the reviewed topic. Furthermore, for the systematically reviewed articles, the average number of citations was 36.32 (ranging from 0 to 206), with Sheng et al. [26] work having the highest number of citations (published in 2020), while for the snowballed articles, the average number of citations was 68.93 (ranging from 5 to 315), with Adedeji and Wang [53] article reaching the highest number of citations (published in 2019). The average number of citations was 52.63 when we consider both systematically reviewed articles and snowballed articles. The most commonly referenced journal was IEEE Access. However, identifying a predominant author was not possible, indicating that there may not yet be dedicated research lines in this field. Instead, the publications appear to be more sporadic.

The five most common general keywords in the examined articles (from most to least common), produced using the pyBibX library (algorithmically generated keywords) on title and abstract, were deep learning, waste management, machine learning, recycling, and artificial intelligence. The frequency of these most common keywords is presented in Figure 3. We can see the top ten general keywords per year of publication in Figure 4, where the prevalence of “deep learning” after 2019 is apparent, along with the emergence of keywords “Computers” in 2016, “sociology” in 2020, and “policy-making” in 2023 reflecting the evolving focus and interdisciplinary nature of research over the years. The graph presented in Figure 5 was produced by adjacency analysis of the general keywords. It is highly relevant to note that “deep learning” and “recycling” (the two nodes with more links) are related to “plastics”. Furthermore, “machine learning” is the third node with more links and is associated with “recycling” directly or through “neural networks”, while the link to “deep learning” by “image classification” and “convolutional neural networks”. The ten most prevalent words (with four or more characters) in the abstract of the examined articles, presented in Figure 6, were (from most to least common) “waste”, “classification”, “image”, “model”, “garbage”, “learning”, “plastic”, “management”, “trash”, and “accuracy”. The most common N-grams with two words were (from most to least common) “deep learning”, “waste management”, “machine learning”, and “artificial intelligence”. From these results, it is possible to conclude that the selected articles are indeed aligned with the scope of the review.

Figure 7 presents a document projection based on the abstract words. The numbers are the references of the articles. The projections were performed using the truncated singular value decomposition method. Embeddings were produced using sentence transformers (all-MiniLM-L6-v2). From the results of Figure 7, it is noticeable that [29] is the most dissimilar, which is expected since it focuses on circular supply chain hierarchical structures, which is a topic that is unrelated to the remaining examined articles. Likewise, we can see [35], which examines enzyme-embedded and microbial degradation methods. [37] is about marine plastic, and [31] also examines the circular economy concept. Reference [67] is also separated as it discusses feature fusion concepts, which are still related to the main research line of other articles but are not prevalent. The big cluster of numbers on the right is associated with the works that are more similar and are aligned with the scope of this review.

3.1. Studies Developed Without Using the Considered ML Models

Huang and Koroteev [27] proposed an integrated approach that uses AI and ML to plan and manage energy and waste, including recycling processes. The study involves the application of neural networks and ML algorithms to predict waste amounts and improve waste collection efficiency. The proposed framework can considerably reduce waste quantities, landfill use, and transportation needs by applying intelligent WM strategies, demonstrating a significant potential impact on improving recycling efficiency and WM. Integrating neural networks for waste prediction and ML algorithms for optimisation in an energy and WM context represents an innovative and practical approach.

Tseng et al. [29] explore the development of a data-driven circular supply chain (CSC). Such a structure is crucial for effective resource management and promoting recycling, both of which are crucial for a cleaner planet. This study highlights the significance of AI and ML in analysing big data to facilitate better decision-making in CSC. The use of data-driven tools such as the fuzzy Delphi method, fuzzy decision-making trial and evaluation laboratory, and entropy weight method demonstrate the application of AI in enhancing recycling and WM practices. This article provides a comprehensive analysis using advanced methodologies to understand and optimise the circular supply chain.

The article presented by Shennib and Schmitt [31] thoroughly examines data-driven technologies and AI applications used in the circular economy and WM systems. The study investigates different applications of AI in WM, such as product lifecycle management, waste generation modelling, community engagement, and waste sorting. These areas play an essential role in advancing recycling technologies and strategies, directly addressing the subject of our research. The article analyses the current state of data-driven technologies and AI in WM, identifying gaps and proposing new areas for research and development.

Yu et al. [33] focus on environmental planning with a focus on the principles of Reduce, Reuse, Recycle, and Recover (4R) supported by an AI-based Hybridised Intelligent Framework (AIHIF). The goal is to promote smarter solutions for a cleaner planet by applying AI and ML in WM and reducing PW. The article proposes a novel AIHIF for WM within the 4R concept, aiming to optimise waste collection, promote recycling, and ensure efficient resource recovery. This approach could considerably enhance recycling rates and overall WM efficiency, contributing to environmental sustainability and cleaner urban management.

Carrera et al. [34] present an economic framework for quality sorting control in plastic recycling classification using ML and spectroscopy technologies. The article incorporates ML algorithms for classifying plastics based on their infrared spectrum, directly addressing the AI and ML aspects. The proposed framework utilises Fourier-transform infrared and near-infrared spectroscopies combined with ML algorithms to classify different types of plastics. The economic analysis of recycling revenue for various polymers and the selection of the most economically advantageous algorithms provide an innovative approach to enhancing the efficiency and profitability of the recycling industry. This could significantly contribute to developing cost-effective recycling strategies.

The article presented by Maraveas et al. [35] comprehensively reviews enzyme-embedded and microbial plastics in agricultural use, focusing on environmentally sustainable solutions. The focus on enzyme-embedded technologies and microbial degradation offers a novel perspective on WM strategies.

Seyyedi et al. [37] provide a detailed overview of the challenges and solutions related to marine plastics in the context of a circular economy, highlighting the crucial role played by AI in achieving a cleaner planet. The report explores the use of AI-based systems for managing ocean PW, explores AI models for predicting the accumulation of ocean PW, and offers insights into policy-making for effective plastic recycling. The article covers a wide range of topics, including the effects of PW on marine ecology, computational methodologies utilising AI, and various approaches to manage and reduce marine plastic pollution.

Kumar and Chimmani [40] examine the application of AI to manage resources and reduce smart home waste. While it may not specifically mention PW, the broader context of waste reduction aligns with this article’s analysis (cleaner solutions for the planet).

Finally, Imran et al. [41] present a comprehensive approach to WM using Quantum Geographic Information Systems (QGIS) for descriptive and predictive data analysis. The article details predictive analytics, a subset of ML, to forecast waste amounts and optimise WM operations. It does not explicitly mention AI or recycling, but the principles and methodologies discussed can be applied to these areas, making it tangentially relevant to the analysis. The use of QGIS for WM is innovative. It reflects a growing trend of incorporating geospatial technologies with ML for environmental solutions.

3.2. Studies Based on the Considered ML Models

ML is constantly advancing, and the success of models for practical applications depends on their accuracy and efficiency. This subchapter provides a detailed analysis of various ML methods, including their detection accuracy, classification accuracy, and combined precision. These methods help to improve areas such as object recognition, detection, semantic segmentation, and instance segmentation.

In waste management, the accuracy of ML methods is essential for waste detection and classification. Waste detection involves identifying waste within a given environment with high precision without false positives. ML models like CNNs and YOLOs are trained to recognise waste objects in complex backgrounds and varying conditions. On the other hand, waste classification categorises identified waste into specific types or materials like plastics, organics, or metals. This task requires a deeper analysis of the detected items, where models like CNNs are further refined to classify the nuanced characteristics of each waste type.

Classification accuracy is also relevant to effective sorting and recycling processes. Detection accuracy focuses on correctly identifying waste items, whereas classification accuracy measures the precision in assigning the correct category to each detected item. Both accuracies address different challenges in the waste management pipeline, with detection serving as the foundational step and classification as the subsequent detailed analysis required for effective sorting and recycling.

Kang et al. [24] highlight the significance of an automatic garbage classification system that employs deep learning techniques. Such systems can be essential for effective WM and reduction, which contributes to developing smarter solutions for a cleaner planet, especially for PW recycling. The article explains how deep learning can automatically classify garbage. The article recommends structural and functional improvements in a deep learning model to enhance garbage classification. These improvements include multi-feature fusion, feature reuse, and optimised activation functions. The article reports high classification accuracy and a quick classification cycle, indicating that this system has the potential to improve recycling efficiency and WM significantly.

Rahman and Das’s [21] research introduces a novel hybrid deep learning framework designed to enhance waste classification, an integral component of effective waste management and recycling processes. By integrating custom-tailored deep learning architectures, including CNNs and EfficientNet models, the study significantly contributes to PW reduction. It leverages advanced AI techniques to accurately categorise various waste types, particularly plastics, underscoring AI and ML’s pivotal role in driving cleaner, more intelligent solutions for environmental sustainability. The proposed methodology aligns with the article’s thematic core and promises substantial accuracy enhancements over existing classification methods.

Alzyoud et al.’s [22] study explores a semi-smart adaptive approach to trash classification, blending physical sorting mechanisms with advanced AI techniques, such as CNNs, to enhance waste management and recycling efficacy. The report presents a comprehensive strategy that resonates with the pursuit of intelligent waste management solutions by incorporating methods such as barcode separation, magnetic separators, and hardness tests alongside CNNs for image classification. The innovative integration of physical and digital sorting technologies exemplifies the practical application of AI and ML in addressing environmental challenges and directly contributes to the goal of PW reduction.

Abdu and Noor’s [23] survey paper examines the application of deep learning technologies in waste detection and classification, offering a broad perspective on the role of AI and ML in enhancing waste management systems. The article reviews image classification and object detection models, showcasing their relevance to recycling efforts and efficient waste management.

Luo et al. [25] introduce an innovative edge-cloud framework that employs Deep CNNs (DCNNs) and YOLOv3 to integrate edge computing with cloud-based services for precise image classification and object detection. The proposed system’s capacity to deliver accurate and rapid detection of recyclable waste while addressing computational and latency challenges inherent in deep learning applications presents a scalable solution with significant potential to revolutionise waste management practices.

Sheng et al. [26] explore a waste management system that synergises IoT technology, LoRa communication, and a TensorFlow-based deep learning model to detect and classify waste items, including plastics. Specifically, TensorFlow allows robust deep learning capabilities in object identification, and LoRa can be used for efficient long-range communication.

Ozdemir et al.’s [5] paper explores applying ML techniques to recycling, including CNNs, support vector machines, decision trees, k-nearest neighbours, and standard artificial neural networks. The paper’s comprehensive analysis provides insights into various ML applications in recycling, demonstrating how these technologies can considerably improve waste management outcomes.

Bhattacharya et al.’s [28] study explores deep learning, particularly CNNs, for automated garbage classification. The introduction of advanced deep learning techniques to improve sorting accuracy represents an important leap forward in minimising the inefficiencies associated with manual sorting methods.

The paper presented by Nafiz et al. [30] introduces ConvoWaste, a novel automatic waste segregation system that harnesses CNNs within a deep learning framework to sort various waste types, including plastics. ConvoWaste exemplifies the practical application of deep learning in recycling, employing advanced AI to differentiate and accurately classify waste materials such as plastics, metals, glass, organic substances, medical, and e-waste into designated categories. Moreover, ConvoWaste innovative use of Capsule-Net for image classification, combined with a hardware setup involving ultrasonic sensors and servo motors for the physical sorting of waste, showcases a holistic and advanced approach to waste segregation. With a reported classification accuracy of 98% and features designed to notify authorities about waste levels, this system presents a significant leap forward in waste management technology.

Wu et al. [32] introduce an intelligent dustbin designed to use ML for effective garbage classification. By employing CNNs for image recognition, the system categorises waste into distinct groups, such as recyclables, kitchen waste, and harmful materials. Integrating advanced technologies, including intelligent speech recognition, sensor applications, and a visual recognition system, into the dustbin design marks a relevant step forward in waste management technology, potentially leading to higher recycling rates. The system’s capability to upload classification data to the cloud further opens the possibility for data analysis and system optimisation.

Altikat et al.’s [15] study examines various DCNN architectures to accurately identify and categorise different waste types. Despite the challenges posed by the inherent properties of PW, such as transparency and deformation, the paper’s approach to tuning DCNN models to improve waste classification accuracy highlights the potential of AI-driven solutions to overcome obstacles in waste segregation.

Tripathi et al. [36] introduce a novel application of the EfficientNet-B3 CNN to waste material classification. With a notable accuracy rate of 97%, the system shows the potential of deep learning algorithms to improve waste segregation.

Zia et al. [38] introduce a reverse vending machine (RVM) designed for efficient PW management, using MobileNet to classify plastic bottles precisely. The RVM’s design, characterised by its affordability, portability, and high classification accuracy, introduces a new, user-friendly approach to waste management. The machine’s success in a university setting, evidenced by substantial PW collection, underscores its effectiveness.

The PLEESE system, introduced by Salim et al. [39], represents a novel intersection of technology and environmental stewardship. It aims to promote the reuse of plastic items by using computer vision and deep learning. By identifying reusable plastic containers through ML algorithms, PLEESE exemplifies the practical application of AI technologies in the promotion of sustainable behaviours. While the system’s primary focus is on the reuse aspect, it indirectly encourages recycling efforts by minimising the volume of waste requiring processing. The innovative nature of PLEESE lies in its strategy to effect behavioural change at the point of disposal, utilising persuasive messaging based on deep learning-driven identification of reusable plastics. This proactive stance on waste management, particularly suited for high-visibility areas like urban centres, has the potential to alter individual behaviours towards plastic use and contributes significantly to the aim of mitigating plastic pollution.

Y. Pan’s [42] study examines domestic garbage classification through the lens of deep learning, presenting methodologies that, while not exclusively focused on PW, are highly pertinent to the broader objectives of enhancing waste management and recycling processes. The exploration of image classification algorithms, particularly through models like ResNet, highlights the application of advanced AI and ML techniques in identifying and categorising waste materials. While the paper’s primary discourse centres around domestic garbage, the implications for PW management are implicit and significant, suggesting that the methodologies discussed could be seamlessly adapted to target PW specifically.

Liang and Gumabay’s [43] study introduces a smart household waste classification system that uses AI to enhance waste sorting and management. While the primary focus is household waste, the AI methodologies employed, including CNNs, hold significant promise for application in general PW management.

The paper presented by Chopde et al. [44] develops an AI-driven system that automatically classifies waste (using computer vision techniques), including plastics. The system operates on a compact and accessible platform, utilising a Raspberry Pi and a camera module, and employs the EfficientDet model for object detection. This methodology demonstrates the feasibility of integrating AI models into everyday waste management tools. It highlights the potential for such technologies to considerably improve the precision and efficiency of waste classification locally. The implications of deploying this technology are substantial, promising to elevate the effectiveness of recycling programs.

Togacar et al.’s [45] study introduces an approach to waste classification that integrates autoencoder networks with feature selection techniques in CNN models, achieving a classification accuracy of 99.95%. This level of performance in distinguishing between various waste types holds immense potential for improving the sorting and recycling of materials, including plastics.

Finally, Ghatkamble et al.’s [46] study introduces an intelligent municipal waste management system that uses the YOLO network alongside IoT technology. This system’s focus is on employing AI techniques for categorising and efficiently managing municipal waste. The innovative fusion of YOLO networks with IoT technology for real-time waste management underscores a novel approach in the domain, offering potential advancements in the efficiency and intelligence of waste handling systems.

3.3. Performance Analysis

In this subsection, we will conduct an analysis of the accuracy of various ML models in object detection and classification. Of the systematically reviewed studies, 16 examined the considered CNN-based models, with their performance being summarised in

Table 2. Overview of the performance of the considered ML models.

ML Models	Detection Accuracy Data Points	Detection Accuracy Average in %	Classification Accuracy Data Points	Classification Accuracy Average in %
Total and weighted average	5	80.11	28	75.36
Snowballed total and average	12	71.36	20	92.05
Global total and average	17	74.86	48	82.62

. The additional snowballed literature and the combination of both (indicated as global) are also included in the table. It is important to notice that the studies used different databases, training conditions, model structures, and number of classes/categories. Thus, this analysis can only be seen as an initial approximation, highlighting the tendencies regarding accuracy in detection and classification. Therefore, it serves as a global overview of the trends. Thus, the rationale is to provide this overview rather than an in-depth examination, which would have required running all models on the same standard dataset.

Regarding the number of categories/classes used by the systematically reviewed articles, it ranged from 2 to 8, with a median of 4 (average of 4.27). More specifically, for detection problems, the number of categories used ranged from 2 to 6, with a median and average of 4. On the other hand, for the classification problems, the number of categories used ranged from 2 to 8, with a median of 4 (average of 4.40). These results suggest the trend of using four categories/classes.

By examining Table 1 and Table 2, it is apparent that the accuracy in detection problems varies substantially, reflecting the difficulties associated with the need to extract spatial features from images effectively. SSD models are optimised for real-time detection tasks, balancing speed and performance adeptly, and can surpass an accuracy of 84%. YOLO models, which reached a detection accuracy of 85%, are known for their efficiency in processing images in a single evaluation pass, providing rapid and accurate detections. It is relevant to notice the global average accuracy of around 75% for the 17 examined samples.

Shifting to classification accuracy, the data highlights the adaptability and effectiveness of CNN in a wide range of image classification tasks, achieving an average global accuracy of around 83%. This reinforces the position of CNNs as versatile tools in ML.

Although SSD primarily focuses on object detection, it also shows potential for classification tasks, with an accuracy of 76.77% [43]. On the other hand, YOLO stands out in classification with an impressive accuracy of 98.05% [46]. However, further research is necessary to validate the capabilities of these models as the representation is from a single article.

Custom models designed for specific applications perform exceptionally well, with an accuracy surpassing 95% [5,23], highlighting the effectiveness of tailored solutions in achieving high performance. Additionally, combining multiple models has proved successful in classification tasks, achieving an accuracy that surpasses 99% [45], emphasising the advantage of collaborative approaches in complex ML challenges.

The convergence of the models’ performance towards a substantially high average detection accuracy is an important milestone in the evolution of detection models, as it showcases their refined ability to interpret complex and diverse visual data landscapes. The same is true for classification accuracy, demonstrating the robustness and adaptability of these models in distinguishing and categorising diverse objects within images. These accuracy-based metrics signify a broader trend towards performance and reliability, driven by the contributions from the examined models, from standard CNN architectures to custom and combined methodologies. YOLO’s notable performance in classification further enriches this narrative, suggesting an expanding horizon for models traditionally associated with detection tasks.

Figure 8 shows a boxplot that summarises the central tendencies and variability of accuracies attained by different ML models. This graphical representation offers a concise yet comprehensive overview of the data, illustrating the distribution and asymmetry of accuracy values and pinpointing any outliers.

Considering the classification accuracy of the systematically reviewed article, it is apparent in Figure 8 that the classification accuracy boxplot displays a wide range of model performance, with the lowest accuracy recorded at 36% [42]. This suggests that some models may struggle with complex detection when considering four classes. There are no minimum outliers, which indicates a consistent performance floor across all models studied. The middle 50% of data points are delineated by the first quartile (Q1, 25th percentile) at 59.24% and the third quartile (Q3, 75th percentile) at 96.28%. This signifies a noteworthy improvement in model performance within this interquartile range, from moderately performing models to those achieving high accuracy. The median accuracy is 77.34%, slightly exceeding the average accuracy (75.36%); thus, the results are reasonably well distributed. The maximum accuracy recorded was 99.60% [38], demonstrating the potential of state-of-the-art models to achieve near-perfect detection in controlled conditions or specific applications. Curiously, the snowballed articles exhibit a much shorter variation, with an outlier at 73.20% [59]. Q1 is 89.80%, and Q3 is 97.11%, with a median of 93.32%. Therefore, in the global classification, reference [42]’s performance is now an outlier; Q1 is 75.40%, and Q3 is 96.33%, with a median of 91.08%.

Regarding the detection performance, it is clear from Figure 8 that a lower interquartile range was attained. For the systematically reviewed articles, Q1 was 65.52%, Q3 was 92.27%, and the median was 85%, which is considerably higher than the mean (80.11%). This suggests a slight left skew in the data, where most models cluster around a higher performance band, with fewer models trailing in the lower accuracy regions. This tendency is even more notorious in the examined snowballed articles, where the median and mean are 79.52% and 71.36, respectively, with Q1 and Q3 of 64.85% and 86.12, respectively. The lowest reported performance, comprising an outlier, was 31.60% [68], while the best performance was 98.30% [30]. Regarding the global detection accuracy, Q1, Q3, and median were 65.70%, 86.62%, and 81%, respectively.

These results highlight a tendency towards good detection and classification performance in the current state of the art. It is especially relevant to the 97.32% classification accuracy attained by Tripathi et al. [36] using eight classes and the 98.30% detection accuracy of ConvoWaste, proposed by Nafiz et al. [30], using six categories. Both articles were published in 2023, showing increased performance even in the more challenging problems with a higher number of categories/classes.

4. Conclusions

This systematic review has comprehensively examined the application of ML techniques in PW detection and classification, revealing relevant advancements in the field over the past five years. Furthermore, a bibliometric analysis was executed to examine relevant keywords, assess which articles are more similar to each other, and assess the citations of the reviewed articles, which is a standard examination that was previously conducted on wastewater treatment with AI [75]. Our analysis demonstrates that CNN-based models, particularly YOLO and SSD architectures, have emerged as highly effective tools in this domain, consistently achieving detection accuracies exceeding 80% and classification accuracies surpassing 83% across diverse studies. Thus, addressing the research questions, it was concluded that ML models achieve a detection accuracy that meets but often exceeds the 80% detection accuracy benchmark. It is also important to notice that in many instances, the accuracy surpassed the 95% threshold. These results support a positive answer to the first research question, “Can ML models achieve PW detection and classification accuracy suitable for real-world applications?”. Regarding the second research question, “Which ML approach is more suitable for PW detection?”, it is more challenging to provide a more direct answer; however, the results of Table 1 suggest that YOLO is likely the best model.

The field has witnessed rapid progress, with recent studies in 2023 reporting classification accuracies of 97.32% for eight waste categories and detection accuracies of 98.30% for six categories, indicating the growing capability of these systems to handle complex, multi-class waste management scenarios. This progress suggests that machine learning technologies are approaching the threshold of practical applicability in real-world waste management contexts.

However, significant challenges remain. The variability in methodologies, datasets, and performance metrics across studies hampers direct comparisons and standardised benchmarking. Furthermore, while laboratory results are promising, there is a notable gap between controlled experimental performance and the robustness required for real-world waste management applications.

These findings have important implications for environmental sustainability efforts and waste management policies. The high accuracies achieved suggest that ML could significantly enhance the efficiency and effectiveness of PW sorting and recycling processes, potentially revolutionising waste management practices. However, successful implementation will require close collaboration between specialised technicians, waste management professionals, and policymakers to address practical deployment challenges. It is important to consider the hardware requirements and the need for interpretability [76].

5. Challenges and Future Research Directions

The recommendation of this review for future research is first to develop a benchmark dataset that allows the examination of all models in the same condition, facilitating direct comparisons between different approaches. Furthermore, it is essential that the used datasets are made publicly available to allow independent validation of the models and results. Lastly, validating the performance and scalability of the proposed solutions in real-world pilot studies is also needed.

Future research lines could examine multimodal approaches by combining image data with other sensor types to improve robustness in varied environmental conditions, explore federated learning techniques to enable collaborative model training, and examine the potential of unsupervised and self-supervised learning methods to reduce reliance on large labelled datasets. It is also advisable to further refine the transfer learning and few-shot learning methods that were used to address the challenge of limited labelled data in specific waste management contexts.

Subsequent reviews should address several key areas identified in this work. As the field matures and more empirical studies become available, a more granular subgroup analysis examining factors such as dataset size, waste category count, and algorithmic variations could provide insights into performance differentials across studies. A comprehensive temporal analysis of model performance evolution would indicate the trajectory and rate of progress within the domain. Furthermore, examining the computational resource requirements associated with various approaches would clarify the feasibility of real-world implementation and provide scalability considerations. From an economic perspective, as these methodologies transition from theoretical constructs to practical applications, rigorous assessments of the economic viability of deploying these technologies across diverse contexts become essential. These assessments would provide valuable data to inform decision-making processes for potential adopters and policymakers.