Solid Waste Generation and Disposal Using Machine Learning Approaches: A Survey of Solutions and Challenges

Abstract

We present a survey of machine learning works that attempt to organize the process flow of waste management in smart cities. Unlike past reviews, we focused on the waste generation and disposal phases in which citizens, households, and municipalities try to eliminate their solid waste by applying intelligent computational models. To this end, we synthesized and reviewed 42 articles published between 2010 and 2021. We retrieved the selected studies from six major academic research databases. Next, we deployed a comprehensive data extraction strategy focusing on the objectives of studies, trends of ML adoption, waste datasets, dependent and independent variables, and AI-ML-DL predictive models of waste generation. Our analysis revealed that most studies estimated waste material classification, amount of generated waste per area, and waste filling levels per location. Demographic data and images of waste type and fill levels are used as features to train the predictive models. Although various studies have widely deployed artificial neural networks (ANN) and convolutional neural networks (CNN) to classify waste, other techniques, such as gradient boosting regression tree (GBRT), have also been utilized. Critical challenges hindering the prediction of solid waste generation and disposal include the scarcity of real-time time series waste datasets, the lack of performance benchmarking tests of the proposed models, the reliability of the analytics models, and the long-term forecasting of waste generation. Our survey concludes with the implications and limitations of the selected models to inspire further research efforts.

1. Introduction

The rate of urbanization and industrialization around the globe is extremely high, and as per the United Nations report, around 65% of the world will be urban by 2050 [1]. On the one hand, this rapid urbanization is indicative of the improvement in the quality of life for the masses; however, it is leading to an array of problems, such as the emission of greenhouse gases, pollution, and municipal solid waste management (MSWM). This relocation to urban areas is directly related to municipal solid waste (MSW) generation. Despite the unprecedented digital revolution, solid waste generation continues to pose severe economic and environmental threats to individuals and governments. In fact, the world currently produces around 2.01 billion tons of solid waste yearly, which is expected to grow to a whopping 3.40 billion tons annually by 2050 [2]. Effective MSWM is mandatory for municipalities, as they usually spend a considerable amount of their budget on managing the waste collection and disposal processes. The effectiveness of MSWM varies from country to country, where typically, low-income countries spend about 20% of their budget while high-income countries spend about 4% of their budget on MSWM-related activities [2]. Mishandling in the MSWM can be costly in terms of environment, economy, and public health [3]. The use of smart technologies has been explored at various levels to support and innovate MSWM processes [4,5,6].

MSWM encompasses several interrelated phases [7]. These phases include waste prevention, generation, storage, collection, transportation, recycling, recovery, and disposal. The waste disposal phase is the most harmful to the environment since it can have a long-lasting impact due to several factors, such as the generation of hazardous gases [6]. Similarly, waste generation marks the beginning of MSWM, and several techniques are applied to minimize waste generation by individuals, industries, and communities. For these reasons, our survey focuses on the tools and techniques related to smart waste generation and disposal behavior and patterns at the citizen and household levels.

The increased waste generation is a natural outcome of urbanization around the globe. Figure 1 highlights the waste generation in different world regions.

As per the estimates, household waste accounts for 55–80% of the total MSW generated from developing countries [6]. On average, each person generates about 0.74 kg of waste annually, whereas North Americans top the list by generating 800 kg of waste per capita per year [8]. Although high-income countries constitute 16% of the world population, their share of global waste generation is 34%. On the other hand, low-income countries produce only 5% of the total global waste generation [9]. Household waste can be categorized into the following types [2]:

• Fruit and vegetable waste;
• Tins, glass, bottles, and plastic bags/containers of waste;
• Paper waste, including newspapers, magazines, etc.;
• Batteries, old medicines, and other hazardous materials such as used oil/kerosene.

Fruits, vegetables, and green waste account for 44% of the total waste generated, whereas dry recyclable waste, including paper, plastic, glass, tins, etc., makes up 38% of the total waste generated globally [2].

Global waste disposal patterns also vary from region to region. A significant 40% of waste is disposed of directly into landfills, whereas only 19% is recycled. Almost 93% of the waste is openly dumped, including burning and dumping by the side of the road in low-income countries; however, just 2% of waste is dumped openly in high-income countries [2] (see Figure 2).

As mentioned above, researchers and the industry have conducted several works to overcome the MSWM challenges efficiently by deploying sensors and smart technologies. Modern solutions are based on the concept of the Internet of Things (IoT), where numerous smart devices are utilized to assist in the different phases of MSWM [4,5,6,10,11]. The deployment of smart waste management technologies paved the way for large data collection throughout the different phases of waste management. Such data can be used to train relevant machine-learning models and techniques [12]. Artificial intelligence (AI) and machine learning (ML) approaches can anticipate waste generation patterns and contribute to efficient waste collection and disposal processes. Indeed, several intelligent solutions, such as [13,14,15,16,17], were originated to facilitate waste management. However, the application of ML in the field of smart waste management is still in its early stage, and further work is required.

Waste management challenges continue to attract considerable research attention, and several reviews exist on smart waste management. However, a limited set of studies reviewed waste management in the context of applying artificial intelligence and machine learning concepts in the field of waste management. This study aimed to cover this gap and present a comprehensive systematic literature review of ML and AI-related applications for smart waste management. We included extensive research studies carried out and published from the year 2010 to 2021. The key contributions of this survey are summarized as follows:

• A detailed analysis related to machine learning and artificial intelligence approaches that assist solid waste disposal and generation activities was presented;
• The survey also focuses on ascertaining factors deemed crucial by researchers to forecast the waste generation and disposal behavior and attitudes of citizens in urban cities;
• We have highlighted significant issues/challenges encountered during the prediction of waste disposal behavior and patterns;
• A comprehensive data collection, extraction, and analysis strategy was devised to filter out the most relevant studies in the field of research;
• Implications and limitations of our survey were also detailed to assist the keen researchers transparently.

The remainder of this survey is organized into six sections. Section 2 presents a literature review of recent waste management works. Section 3 details the methodology that was adopted in this survey. Section 4 presents the key results of our analysis. Section 5 discusses the main findings, challenges, and limitations of our survey. Finally, Section 6 concludes our research and highlights future research directions.

2. Related Works

This section sheds light on the previous works concerning smart waste management solutions and waste disposal behavior. Moreover, it concludes by contrasting the existing surveys on SWM and highlighting the research gaps within the previous SLRs.

2.1. Smart Waste Management

Waste management refers to the complete cycle of treating waste through the provision of practical solutions to each stage in the cycle [18,19]. The term “waste management” covers all types of waste that are generated by various segments of the community, irrespective of the composition and state of waste [20]. Solid waste management (SWM) is considered one of the waste types that can cause significant adverse environmental issues [21]. For over a century, the field of waste management has gained enormous interest due to many reasons. It is an active area of research where many researchers and companies focus on waste analysis and how to utilize this analysis in real-world scenarios. Advances in the field led to deploying smart waste solutions by adding intelligence to waste-related infrastructures, such as bins and trucks [22]. As a consequence, various advanced monitoring technologies have been developed, such as geographic information systems (i.e., GISs), ultrasonic sensors, radio-frequency identifications (RFID), etc., [21].

Smart waste management can be defined as deploying and using smart technologies for more effective, productive, and sustainable operations in waste management [6]. These technologies are developed based on various computing fields, such as artificial intelligence, IoT, cloud computing, fog computing, and big data. These fields are applied in different stages of the waste management cycle. Typically, this cycle encompasses the following intertwined stages: disposal from bins, source separation, collection, waste treatment, and transport final disposal [23]. Other works defined the waste cycle as a waste collection system together with its transportation, disposal, and recycling [20].

The Internet of Things (IoT) is gaining traction in various industries, including SWM. IoT-based SWM for smart cities not only enhances each phase of the SWM cycle [15,16] but also allows for the implementation of eco-friendly SWM [18], which has a significant impact on the circular economy [17]. Smart-bin technology [19] based on IoT allows us to track real-time location and filling status, which aids them in collection decisions [21], monitoring temperature, identifying types of waste material, and sending instant alerts in the event of an emergency. The purpose of this paper was to review the existing ML-driven solutions and challenges related to solid waste generation and disposal.

2.2. Advantages of Waste Disposal Solutions

Several technologies were utilized in waste disposal solutions. The Smart-M3 platform was deployed to intelligently monitor and predict when to collect solid waste bins [24]. The solution took advantage of IoT technologies by installing filling sensors within garbage bins. As a result, the approach empowered truck drivers to optimize their routes and save fuel. Another solution called Smartbin system was implemented based on the wireless mesh network [18]. The approach helps cleaning operators plan their waste collection trips. Moreover, an innovative cloud-based smart management solution was implemented to manage waste [25]. The solution provides several advantages, including timely waste collection, route optimization, recycling, and disposal. Furthermore, an IoT-based approach was presented and deployed for smart garbage monitoring and clearance in India [26]. The solution uses a microcontroller as a middle layer between the sensors and a GSM/GPRS system.

An IoT-based infrastructure applied to a waste management scenario was installed to manage waste in smart cities [27]. The study deployed three protocols, namely CoAP, HTTPS, and MQTT, for secure communications. The CoAP-based approach was concluded as the best-performing solution with regard to the number of concurrent operations, cost, and quality of service. Moreover, an Industry 4.0-based solution was implemented to tackle waste management while achieving sustainable development goals [28]. The approach utilized industry 4.0 technologies for smart and sustainable waste management, including the disposal stage. In addition, a multi-layer long-range wide-area network approach was utilized for smart waste management [28]. The solution encompassed video surveillance units, which are fed by machine learning capabilities for more smart city features. It provides capabilities to observe and manage bins and trucks’ filling levels while keeping track of their states. Finally, Blockchain technology was deployed to manage and trace solid waste [29]. The approach showed promising results in terms of data tracking, data sharing, and waste control. It is also notable to mention that the cost of this approach is manageable and inexpensive. All of these solutions aimed to forecast and optimize the waste disposable behavior and patterns for a more sustainable future.

2.3. Waste Disposal Behaviors and Attitudes

A multitude of smart models was deployed as part of waste disposal solutions. A gradient-boosting regression model was implemented to predict short-term waste generation across New York City [30]. The model achieved 88% average accuracy due to the fine temporal and spatial granularity of the optioned dataset. The result of waste generation was aggregated to one-week timescales due to the regularity of waste behaviors. The model was able to observe fluctuations related to holidays, seasonal events, and gatherings. Furthermore, the paper discussed the presence of external factors such as the weather, where the amount of waste disposal was impacted by extreme weather events, such as hurricanes and snowstorms. However, a corresponding makeup collection must be prepared shortly after to avoid the buildup of unpicked waste. A deep learning strategy was applied to predict multi-site household waste disposal [31]. The study used a multi-site long short-term memory neural network (LSTM) to forecast waste disposal in the municipality of Herning, Denmark. The study results showed that the multi-site approach achieved better forecasting compared to the benchmark models. On average, households produce between 30 and 80 kgs of waste each week. This study, in contrast to the previous one, showed no significant correlation between weather factors and waste disposal behaviors. Another study employed LSTM to predict and monitor waste generation in smart cities [31]. The study collected temperature and humidity readings which can be used to calculate the carbon monoxide levels generated by the waste.

An innovative waste management solution based on a convolutional neural network algorithm was developed to classify the waste type and predict future disposal patterns [32]. The approach also supported assigning a price to recyclable items inside the Automated Teller Dustbin (ATD). The model achieved an average accuracy rate of 96% in recognizing various objects thrown in the ATD. Moreover, a solution based on artificial neural networks (ANN) and decision tree algorithms was created to predict and monitor municipal waste disposal in Canada [33]. The study showed that the neural network models had a better prediction rate, estimating around 72% of the variation in the data. Another study focused on predicting waste disposal over a municipal level by combing a support vector regression model and genetic algorithm (GA-SVR) [34]. The solution, implemented in Huangshi, China, could predict waste disposal over intervals and has a high accuracy rate in generalization ability. Finally, gradient-boosting regression trees and neural network models were implemented to calculate waste disposal in New York City (USA) [35]. The solution calculated the daily and weekly waste disposal for around 750,000 properties. The approach can forecast building-level waste disposal with a high-performance result.

2.4. Waste Management during COVID-19

The COVID-19 pandemic has dramatically influenced all aspects of life, including disposal behavior and patterns, so machine learning-based smart waste management solutions must adapt to the new conditions. While there are various studies discussing medical waste [36,37,38,39], there is a lack of research investigating the impact of COVID-19 on smart waste management systems. Nonetheless, a few attempts to tackle the problem utilize more conventional methods. Researchers proposed a blockchain-based solution to tackle supply chain and estate management during the COVID-19 pandemic [40]. Furthermore, the authors in [41] proposed an intelligent solution based on IoT principles to treat medical and home waste. More research is required to investigate the impact of COVID-19 on waste management systems, mainly when using ML-based architectures.

2.5. Existing Surveys

In order to clarify the motivations for conducting our literature review, we analyzed the literature with respect to the recent surveys in the selected area of research, as shown in

Table 1. A Comparison of smart waste management surveys (S = survey, LR = literature review, J = journal, C = conference, Y = yes, N = no).

REF	Year published	Survey type	Publisher	Venue (J/C)	Years covered	Papers reviewed	Quality Assessment	Focus of Survey	Machine Learning	Research Issues	Strengths	Weaknesses
[42]	2020	S	Elsevier	J	2004–2019	85	Y	AI applications in waste management	AI applications	Y	(+) This review highlights the importance and needs for adopting AI techniques in the SWM	(−) AI techniques depend on secondary data to train and adapt themself in such cases to successfully run this model. The need for reliable datasets
[43]	2018	S	IEEE	C	2016–2017	10	N	Prototype-based technology adoption in SWM	N	Y	(+) Describes the importance and usage of low-cost sensors in SWM	(−) No focus on the usage of machine-learning techniques
[44]	2016	LR	Elsevier	J	2010–2013	87	Y	Focuses on the absence of planning and infrastructure for the disposal of solid waste	N	Y	(+) Discusses decision support systems (DSSs) theories to assist authorities in complying with SWM regulations	(−) Mainly focus on the current in-hand issues and their management while little focus on future scenarios is considered
[45]	2017	S	IEEE	J	2003–2015	32	Y	Focuses on intelligent waste collection (transportation) using IOT sensors	N	Y	(+) This paper mainly categorizes SWM into three categories and focuses on improving the waste collection process	(−) This survey is restricted to a single city (St. Petersburg, Russia) and emphasis on the routing problem of waste collection
[20]	2018	LR	MDPI	J	2014–2018	15	Y	Focuses on the solid waste categorization based on its material and also focuses on describing the technical architecture of the different layers in IoT sensors/microprocessor	N	Y	(+) This review contributes to categorizing solid waste into different material levels and deeply describes the architecture of the IoT sensors	(−) Too many technical details are provided in terms of IoT sensors architecture rather than planning and its usage
[11]	2018	LR	Elsevier	J	2008–2017	17	N	Focuses on product’s life cycle and introduces a new business model in the process of product life cycle	N	Y	This paper introduces a sharing economy that enables citizens to declare their possession of the product as either supply or demand through an online portal, which helps the SWM to foresee the waste generation rate	This new sharing economy model will only work when the citizens are responsible enough to provide such data, which is very hard to accomplish.
[46]	2019	LR	Elsevier	J	2001–2019	85	N	Focuses on the adoption of robotics technology in SWM	N	Y	This paper focuses on the robotics-based sorting systems in SWM that will reduce human involvement in a very efficient, timely way	The proposed robotics technology is very costly and also requires a lot of hardware replacement and maintenance, which is not discussed
[47]	2015	LR	Elsevier	J	1996–2015	98	Y	Focuses on ICTs and their application in SWM	N	Y	It broadly categorizes ICTs into four categories, i.e., spatial, identification, acquisition, and data communication. It also highlights SWM systems made from these the combination of these four categories	This paper highlights how ICTs and communication work in different scenarios.

.

3. Methodology of Our Literature Survey

This section explains the procedure that we followed in our systematic literature review. It comprises five subsections, namely motivating research questions, search strategy, search process and inclusion criteria, data extraction, and quality assessment.

We adopted the standard literature review methodology advocated in renowned publications in software engineering, precisely that described in [48]. In essence, a systematic review endeavors to synthesize and understand the existing works published in different research venues in an organized and succinct manner [48]. To this end, we started the review process by exploring the existing reviews and surveys that investigated the use of computational intelligence for smart solid waste management (SWM). The initial overview helped us to (1) identify the gaps in current surveys and (2) scope the research questions to motivate our survey. Typically, a literature review follows four subsequent steps: establishing a solid search strategy, executing the search protocol, analyzing the review results, and finally compiling the results [48]. In our search strategy, we utilized a selection of keywords that are most relevant to our topic of choice in six renowned electronic databases. Next, the most relevant studies from these databases were identified, where the required data were extracted accordingly. Finally, we compiled and presented the key findings in view of the survey objectives. Typically, a survey protocol can be divided into five phases, as advocated by previous guidelines [48,49], with some variations between the studies. Below we explained each phase.

3.1. Motivating Research Questions

In summary, our review aimed to combine and clarify the empirical evidence concerning the ML-based waste management models, which are developed to predict waste disposal and generation activities. In order to extend our understanding of the computationally intelligent approaches applied in this field, we also included artificial intelligence and deep learning models. Three fundamental research questions (RQs) were posited to motivate this survey as follows:

• Research Question One (RQ1): What are the most used machine learning approaches and algorithms developed to optimize solid waste disposal and generation activities?
• Research Question Two (RQ2): What independent variables (features) are used to forecast the waste disposal behavior and attitudes of citizens and urban cities?
• Research Question Three (RQ3): What are the major issues/challenges encountered during the prediction of waste disposal behavior and patterns?

3.2. Search Strategy

In order to identify the relevant articles in our selected field, we explored six popular academic research databases, namely (1) IEEE Xplore, (2) ACM Digital Library, (3) ScienceDirect, (4) Web of Science, (5) Springer, and (6) Google Scholar. We focused our quest on the research articles published between January 2010 and December 2021. After reviewing the technology trends and central areas of smart waste management, we specified 19 search keywords and clustered them into two categories, i.e., Technology and SWM, as listed in

Table 2. Keywords Used to Guide our Search (C = compulsory, O = optional).

Technology Keywords	Smart Waste Management (SWM) Keywords
Artificial Intelligence (C)	Dispose of waste
Machine Learning (C)	Disposal of waste
Deep Learning (C)	Waste disposal behavior
IoT Technology (O)	Waste disposal attitude
Robotic Technology (O)	Household refuse
Image Recognition/Classification (O)	Household waste disposal
	Disposing of household waste
	Waste disposal practice
	Household
	Individual
	Citizens

The above search key terms were combined using Boolean operators (AND, OR) to construct complex search phrases to help retrieve candidate articles from the selected databases. The syntax of the search phrases had to be slightly adjusted to accommodate the variability of the databases.

. It is important to note that the key terms were selected systematically. The search terms were highly linked to the stipulated research questions.

3.3. Search Process and Inclusion Criteria

A comprehensive search was performed on all six database resources through the application of the subsequent stages:

• Initial search: first, a random search using a fuse of the chosen keywords and patterns was performed on the databases mentioned above without applying any exclusion criteria;
• Application of inclusion criteria: in this phase, the obtained articles were filtered by applying several inclusion criteria so to keep the relevant research works:
  • I1: the study must appear in either a peer-refereed Journal or Conference venue;
  • I2: the study must be peer-reviewed by the scientific community;
  • I3: The study must be conducted by the authors (i.e., primary studies);
  • I4: the study must be published in the year range of January 2010 to December 2021;
  • I5: the study must be written in English;
  • I6: the study must discuss the use of artificial intelligence, machine learning, or deep learning in predicting solid waste disposal behavior or attitudes.
• Application of exclusion criteria: next, we reduced the subset of candidate studies by eliminating the articles that fall within these categories:
  • E1: literature review or survey papers (i.e., secondary studies);
  • E2: short articles (less than five pages);
  • E3: the grey literature (preprints, unpublished works, and technical reports).
  • E4: duplicate articles from the same authors; in which case only the most recent and extended article is selected;
  • E5: papers that are irrelevant to the main contributions of our survey; that is to say, articles that focus their discussions on other solid waste management phases, such as waste recycling, energy conversion, and biological treatment;
  • E6: papers that do not present artificial intelligence or machine learning approaches or solutions.
• Final selection: a fourth filter was applied to the retrieved articles by reading the titles, abstracts, and full text to determine the eligibility of the candidate papers that assemble answers to our solid waste disposal research questions.

Our initial search resulted in 370 possible articles. After carrying out the above phases and applying the proposed criteria, the final list of papers was reduced to 42 relevant studies.

3.4. Data Extraction

Relevant data were extracted from the selected articles to answer the research questions postulated in Section 3.1. Therefore, we constructed a data extraction form consisting of five sheets to capture diverse aspects of smart solid waste disposal behavior and patterns. The data extraction and analysis form includes the following sections:

• General information: this tab captures details about the publisher, publication year, type, venue, and discipline;
• The article’s focus: this tab summarizes the paper’s aim along with the key findings;
• Dataset: this tab details the dataset (whether primary or secondary) used in each article. It also lists how these data were collected and other model performance parameters;
• AI-DL-ML solutions: this tab represents the artificial intelligence, deep learning, and machine learning techniques used in each study and their benchmarking details against the other techniques;
• Quality assessments: this tab summarizes the quality of the selected articles.

Once the data extraction process was completed, the key findings were summarized and highlighted regarding several aspects of the retrieved data to fulfill the objectives of this review.

3.5. Quality Assessment

The last step of the search protocol focused on performing an appraisal of the smart waste disposal studies. The quality of articles was judged by scoring a set of relevant criteria that assess various aspects of the study.

4. Results

This section discusses the different aspects of the surveyed articles in terms of solid waste management (SWM) using machine learning (ML) techniques. The results section is divided into four sub-sections, i.e., publication venues and years, research focus, datasets used, and proposed AI-ML-DL solutions.

4.1. Publication Venues and Years

A total of 42 studies were considered and analyzed in the presented survey. Figure 3 shows that smart waste management solutions were presented in peer-reviewed journals (40%) and conferences (60%). ML is a revolutionary computational paradigm that is currently applied to many fields. For SWM, we concentrated on the last eleven years of research to study the progress of AI/ML/DL techniques in waste generation and disposal.

Figure 4 shows the overall adoption trends of ML in SWM and depicts that research interest rose to its peak in the year 2020. It is clear that the last three years (i.e., 2019–2021) witnessed a considerable rise in the SWM research works, representing 83% of the total articles reviewed.

It can be seen in Figure 5 that the best quality sources have more weightage in terms of content retrieval. The priority of sources is reflected by the percentage of papers surveyed from them. Approximately 67% of the papers surveyed are indexed in IEEE. Similarly, papers indexed in Science Direct, ACM, and Elsevier were also considered. The percentage of papers from each of these indexation services is reflected pictorially.

Figure 6 shows the number of articles searched in various disciplines, including computer science, energy, geography, public health, etc. The number of papers from which different contents are collected is visually presented in front of each discipline. Computer science, multidisciplinary, environmental sciences, energy, and geography constitute the biggest number of articles, i.e., 22, 7, 4, 3 ([50,51,52]), and 2 ([35,53]), respectively.

Figure 7 shows the frequency map of the different keywords in the domain literature considered for this study, with machine learning, waste management, and artificial network as the most used terminologies by the authors. The big spheres depict the most repeated keywords, while smaller spheres represent less repeated terms. Consequently, it can be seen that machine learning, waste management, and artificial neural network are the most repeated keywords found in the literature. Similarly, the small spheres depict less frequency of the corresponding keywords.

Figure 8 demonstrates the venues of publications, shown on the y-axis, and the number of articles retrieved, shown on the x-axis. There are three papers published in the IEEE Access journal [54,55,56] and two papers presented in IEEE International Conference Humanoid, Nanotechnology, Information Technol Communication Control Environment Management (HNICEM) [57,58]. The rest of the papers are published in well-reputed journals and conferences.

4.2. Focus of the Research

This survey considered publications of 24 city-wise studies and 36 country-wise studies. The details of the number of publications related to various cities are purported in Figure 9. It is worth mentioning that 18 out of 42 publications did not mention details of their cities. Similarly, Figure 10 presents the country-wide details of publications, depicting that 13 out of 42 studies were conducted in India, while six (i.e., [51,57,59,60,61,62] articles did not report details of their country. The majority of studies were conducted in Asian countries.

Figure 11 shows that, out of 42 publications, 34 (81%) studies considered only one variable as their primary research outcome. At the same time, only four (9%) studies predicted two variables [63,64]; one study predicted three variables [35], one study predicted four variables [58], one study predicted nine variables, and one study predicted 40 variables. Material classification, waste generation per area, waste filling level, and other variable utilization are characterized in Figure 12. It can be observed from the figure that waste material classification is investigated in 19 (45%) research papers, while waste generation per area and waste filling level predicted variables are used in 8 and 11 studies, respectively. This survey divided the selected articles into seven general categories based on their problem identification and prediction, i.e., waste management areas or stages. In such a way, Figure 13 portrays the number of articles belonging to various waste management areas or stages, revealing that 14, 7, and 8 publications examined waste segregation, waste collection, and waste generation, respectively.

4.3. Waste Management Datasets

All final articles used at least one waste dataset to assess the performance of their prediction models. In this survey, we thoroughly analyzed and extracted various aspects of these datasets to assist us in answering the questions posited in our research. Figure 14 shows that 60% of the datasets used in the articles are publicly available and can be re-used for further investigation, while 40% of the datasets are private and are only limited to the specific article implementation.

Figure 15 shows the number of datasets utilized by various applications. Thus, 35 papers out of 42 (83%) utilized only one dataset, while 3 papers utilized two datasets [64,65,66], 2 papers utilized four datasets, 1 paper utilized seven datasets, and 1 article did not utilize any [67].

Figure 16 portrays the focus of each waste dataset. In this review, we categorized the datasets into ten possible phases. Thus, 11 (26.19%) research articles addressed waste generation, while 9 (21.42%) publications relate to waste material segregation, 5 (11.90%) publications related to waste collection [17,58,60,68,69], and 4 (9.5%) studies relate to waste material classification [56,70,71,72,73]. The number of other publications concerning each waste management phase is presented in Figure 16.

Figure 17 shows the percentage of the training and real-time datasets considered in the selected studies. In training datasets, data were collected from secondary sources, refined, processed, and studied, and based on the analysis, and a prediction model was proposed and implemented. However, in the real-time dataset, a prediction model is selected in advance, and then a suitable real-time dataset is generated in the laboratory and passed to the prediction model. Figure 17 shows that 88% of the models are trained and evaluated over a training dataset, while 7% used a real-time dataset.

Figure 18 presents details of different IoT sensors used by the selected studies. These sensor data are either provided by public entities, such as trashNet [60] or collected from the prototypes developed and implemented for collecting the datasets. It can be observed from Figure 18 that microcontrollers and cameras are utilized in 12 studies, while ultrasonic sensors are utilized in 8 studies. Details of other IoT sensors are depicted accordingly in the below figure.

Figure 19 shows the range of the dataset sizes used in the 42 selected articles. It can be observed from the figure that 16 (38%) out of 42 papers did not specify the size of their datasets. Six (14.28%) articles utilized datasets of less than or equal to 1000 records [13,51,62,69,74,75], nine (21.42%) articles used datasets ranging from 1000 to 5000 records, [33,60,61,76,77,78,79,80,81], and one article used a dataset size greater than 5,0000 data points.

Figure 20 depicts the intervals at which waste data were collected or generated in our selected studies. Thirteen (30.95%) studies recorded waste data observations daily, three (7.14%) studies yearly [33,50,82], two (4.76%) studies monthly [67,83], three (7.14%) studies weekly, and one study quarterly [52]. Notably, 20 (47.61%) research papers did not mention the timeline for generating their datasets. Furthermore, dataset generation time is categorized as a time series or non-time series. Figure 21 shows that 56% of the datasets are time series where data points are recorded over a uniform time interval, and non-time series data are generated arbitrarily or in a non-sequential manner.

Figure 22 presents the ratio of training and test datasets utilized by different prediction models. The details of test and training datasets are presented for each paper. For example, Dessai et al., 2021 utilized 80 percent of the dataset for training the model [77] and the remaining 20 percent for evaluating the learning model’s performance. However, Figure 23 specifies the different independent variables (x-axis) and their general category (y-axis). Eighteen (42.85%) models used images of waste type, ten (23.80%) models employed demographic data, and five (11.90%) models used images of waste fill level [17,69,72,84,85] to produce their predictions.

Figure 24 elaborates that 79% of the data relates to demographics, which includes the nature of a population region such as age, job, location, and family members, 14% relates to the socio-economic factors [35,51,65,67,83], 5% to both preceding clauses [33,53], and the remaining 2% of the data belongs to sensor-related data [78].

4.4. AI-DL-ML Solutions

This section details the machine learning models opted for the automation of SWM by the studies analyzed in this survey. It also highlights other algorithms that are used as a benchmark to compare and select the best performance results. Figure 25 presents the number of articles deploying various kinds of ML algorithms. It can be observed that 15 studies (35.71%) have deployed neural network-based algorithms, while four (9.52%) studies have deployed K-nearest neighbor algorithms ([13,52,77,78]). Figure 26 presents the prediction algorithms compared against the domain literature. It is demonstrated that random forest, decision tree, and artificial neural network classifiers were used as benchmark algorithms in four studies [54,77,83]. Notably, 18 papers (42.85%) did not report any comparison results of their developed algorithms, which questions their validity.

Figure 27 provides details of the prediction model duration, where it is observed that 15 (35.71%) research works provided waste predictions daily, five research papers provided hourly and monthly predictions [52,61,69,78], and three research papers [33,51,53], predicted waste yearly. Twelve studies did not provide details of the prediction model frequency.

Figure 28 depicts the ranking of journal papers in Q1, Q2, Q3, and those papers not published under the Q category. As per the analysis, 26% of the 42 research papers are published in Q1 category [17,35,54,56,65,67], 7% are published in Q2 [52,63,82], and 3% of the papers are published in Q3 journals [53]. The remaining 64% of the studies are published in journals with no recognizable ranking.

Figure 29 shows that 26 research papers deployed multi-step forecasting, while 16 deployed a single-step forecasting technique. Figure 30 shows that 23 out of 42 papers produced better results than the other benchmarked algorithms. A total of 41 research papers employed supervised learning, while the remaining research paper implemented an unsupervised machine learning approach [69]. Figure 31 details the number of research papers providing time series prediction. In time-series prediction, articles predict SWM in different sections, i.e., waste generation, classification, finding the best route for waste collection, etc., over an order period. There are 30 research papers wherein time series prediction is provided, while the remaining 12 do not provide such a prediction.

Figure 32 highlights the overall quality score of the selected studies. All studies were subjected to a thorough assessment of their quality against 14 criteria, as listed below. Each criterion was assigned a maximum of one mark depending on the fulfillment of the criterion (CR).

• CR1: Clear research contributions;
• CR2: Well-defined research questions;
• CR3: Sound research methodology;
• CR4: Datasets described in detail;
• CR5: Predictions (i.e., dependent variables) made are clear;
• CR6: Independent variables (i.e., features) clearly described;
• CR7: Separate training and testing datasets;
• CR8: Validation/ verification method detailed;
• CR9: Detailed analysis (comparison against baseline models) performed;
• CR10: Verification of predictive model with a second dataset;
• CR11: Empirical results detailed sufficiently;
• CR12: Threats to validity discussed;
• CR13: Research implications (theoretical, practical) suggested;
• CR14: Research limitations/challenges discussed;

As per our drafted quality criteria, Rutqvist et al., 2019, [17] scored the highest score, 14. Similarly, Kannangara et al., 2017, [33] scored 13 points. The study with the lowest score is Sudha et al., 2016, [72]. The scores of the other selected studies can be viewed in Figure 32.

Figure 32. Overall quality evaluation of the selected studies.

Figure 32. Overall quality evaluation of the selected studies.

5. Discussion

This section reports on the main findings from the surveyed papers, presents the challenges involved in machine learning-based waste management techniques, and summarizes the limitations of current work and directions for future work.

5.1. Main Findings and Implications

Our analysis is based on 42 selected articles, where the focus of the studies was on the application of computational intelligence techniques (i.e., AI-ML-DL) for solid waste management (SWM). Below we attempted to answer our research questions. The first question (RQ1) enquires about the most used machine learning approaches and algorithms developed to optimize solid waste disposal and generation activities. Most of the surveyed papers either predict the quantity of waste generation or classify the type of waste for appropriate and sustainable disposal. Meza et al., 2019, [67] proposed to predict waste generation based on the previous month’s recorded data and achieved the best results using a support vector machine (SVM) model. Hussain et al., 2020, [52] predicted air pollution caused by improper solid waste management (SWM) processes. They found that the K-nearest neighbor (KNN) algorithm produced the optimal results on real-time data compared to other techniques (e.g., logistic regression, multilayer perceptron, Naïve Bayes). Adeyemo et al., 2016, [78] used a neural network model to estimate the bin status to facilitate waste collection decision strategies. Behera et al., 2020, [51] proposed a model to classify waste materials into two classes visually, i.e., biodegradable and non-biodegradable, using the YOLOv3 algorithm along with SVM. Corpuz et al., 2018, [58] used fuzzy logic control and different sensors to achieve robotic indoor house cleaning. Dessai et al., 2021, [77] demonstrated that K-nearest neighbor classifiers produce better classification results than random forest classifiers when classifying waste material into biodegradable and non-biodegradable waste classes. Farinella et al., 2019, [84] developed a prototype to estimate the amount of food waste inside the dining hall using a CNN-based image recognition model. Fonseca et al., 2019, [69] found that CNN’s “Inception_v4” model achieved higher accuracy than “ResNet50” and “Darknet53”. Hua et al., 2020, [61] proposed scanning real-time video streams through OpenCV and CNN, and their proposed prediction model identified hazardous material with a 90% accuracy. The model proposed by Savla et al., 2020, [64] predicted waste-bin status with an 80.36% accuracy. Shamin et al., 2019, [82] successfully segregated degradable and nondegradable wastes. Adeyemo et al., 2019, [74] used a multilayer perceptron artificial neural network to predict waste bin status with 98% accuracy. Sirawattananon et al., 2021, [71] used Restnet-50 (CNN) to classify images of the waste with a 98.8% of accuracy. Khoa et al., 2020, [63] proposed a model that employs the sigmoid function to predict the probability of waste collection. Dubey et al., 2020, [13] created a KNN-based model that predicts the waste material with a 93.3% accuracy and alerts the municipal authority of the current status. Kontokosta et al., 2018, [35] used a gradient boosting regression tree (GBRT) to achieve a 99.8% prediction accuracy, which is higher than a simple neural network model. The smart model by Baby et al., 2017, [68] used SMS and Email to alert the respective authority to collect a full waste bin, which helped municipalities to adjust and optimize their transportation routes. Yang et al., 2021, [73] proposed weekly supervised learning to improve garbage classification, especially in the case of insufficient waste data.

Overall, CNN achieved the highest accuracy for classifying waste in most studies. The VGG16 model for transfer learning gave the best results among the CNN models. Most of our articles utilized a dataset for model training and prediction. However, only 57 percent of the dataset used in these articles are available publicly and can be re-used for further investigation. Eighty-eight percent of the studies utilized a training dataset. Overall, only 30 research papers provide time-series predictions.

Our second question investigated the independent variables (features) used to forecast the waste disposal behavior and attitudes of citizens and urban cities. Approximately 70% of the developed models used no more than three features to predict waste generation and disposal. The three features pertaining to the provision of future waste estimations included images of waste type, images of fill level, and sensed bin fill levels. Other studies used demographic data to estimate waste generation. For instance, Ahmad et al. [54] considered various independent variables in each grid population (e.g., age, gender, weekdays, etc.) and applied the SVM to predict waste generation.

5.2. Research Challenges in Smart Waste Management

The studies investigated in this survey highlighted some critical issues with machine learning-based waste management solutions. The challenges of waste generation and disposal constitute the answer to our third research question (RQ3). The first challenge concerns the reliability of the forecasting models and the accuracy of solid waste predictions. Inherently the performance of machine learning models depends mainly on the quality of datasets; thus, waste disposal predictions are highly impacted by the availability of waste data and the dataset collection process. There were cases where faulty sensors led to incorrect data [52]. Moreover, generalizing predictive models to multiple cities requires further investigation due to their intrinsic variances in population characteristics and infrastructure.

The second research challenge is caused by the lack of historical data, which degrades the performance of time series-based regression models for predicting the waste amount disposed of in urban cities. Similarly, for the visual classification of waste categories, the performance is significantly degraded by incomplete or damaged waste images [59]. Moreover, environmental problems pose numerous difficulties and obstacles for real-world automated waste collection activities [58]. We found that most models used one modest dataset to evaluate their prediction power, which is considered a major limitation.

The third challenge is linked to the need for a dedicated and resourceful IoT infrastructure. Such hardware implementations are costly and unattractive within smart cities. The heavy dependency on sensing devices or smartphones restricts the wide-range implementation of smart waste management models. Until these disruptive IoT technologies become seamlessly integrated (e.g., plug and play) within the smart waste infrastructure, managing waste processes using intelligent computational models will remain a significant issue.

Indeed, ANN provides foundational concepts of learning algorithms in machine learning; however, optimization techniques are required to enhance the performance of ANN variants. Scientifically it is not easy to choose the appropriate ML algorithm under different circumstances and settings [67,86], demonstrating the fourth challenge.

5.3. Limitations of Study

It is well known that systematic literature reviews help attain a solid understanding of a specific research domain by addressing focused research questions. In our SLR, we applied survey standards and guidelines to search and analyze various academic databases. However, as with any other research study, this SLR also has some limitations. Firstly, to find out the most relevant research articles, we carried out both automated and manual searches by varying the search key phrases. Even though we attempted to include every possible relevant study, we were constrained by the quality of our devised research criteria. Secondly, we limited our search to only six electronic bibliographic databases. Therefore, other academic databases, such as Tylor & Francis, were not searched. As such, we might have missed articles from other databases. Thirdly, we restricted our search to models that were published between the years 2010 and 2021. Hence, there is a possibility that our analysis might not have included some essential predictive models published before 2010 or after 2021.

Some recent efforts by the research community focused on developing intelligent waste management solutions that promote awareness and environmentally responsible behavior. For instance, authors in [87] discussed the possibility of designing a smart bin based on human emotions and experiences. Additionally, some published works presented the latest trends in smart waste management, such as using electronic noses for analytical measurements [88]. The notion of circular economy is also under the spotlight, where researchers are exploring efficient solutions, practices, and policies that would lead toward sustainable MSWM. Researchers are proposing to utilize technology, governance, and industrial support to develop such solutions [89,90]. Illegal waste dumping and disposal is another challenge that is being pondered upon by the research community. Illegal waste dumping poses a severe danger to the environment and pivotal natural resources such as underground water. Efficient waste management policies are required to discourage such behavior [91,92]. These studies were excluded from our analysis since they do not directly present AI and ML-based architectures; moreover, these studies were published after 2021. Nonetheless, these recent efforts set the precedence for future research directions in the field of smart waste management. Moreover, the selected studies created predictive models for household waste instead of commercial waste. However, commercial waste has unique characteristics and should be investigated in separate research works.

We want to emphasize that despite these limitations of our SLR, (1) we explored the major academic databases, and (2) we defined a clear and comprehensive inclusion criterion so that we could find most of the relevant articles.

6. Conclusions and Future Directions

In this systematic literature review (SLR), we surveyed studies that present predictive models and algorithms for optimizing and managing solid waste generation and disposal processes. This research topic is of utmost importance since municipal and household solid waste is increasing at an alarming rate. Moreover, smart waste management has become a necessity for municipalities to reduce their spending and optimize their resources while running waste management processes efficiently. The proliferation of smart technologies, led by artificial intelligence (AI) and machine learning (ML), has paved the way for developing data-driven smart waste management solutions.

In this survey, we synthesized and analyzed a total of 42 distinguished studies that implemented artificial intelligence and machine learning approaches to oversee and manage solid waste generation and disposal effectively. All of our final studies were carried out and published between the years 2010 and 2021. The studies were selected by applying rigorous inclusion criteria to candidate articles retrieved from IEEE Xplore, Web of Science, Springer Link, ACM Library, Science Direct, and Google Scholar. Moreover, our SLR explored fundamental factors that are deemed crucial by researchers to forecast the waste disposal behavior and attitudes of citizens in urban cities. After deploying a comprehensive data collection, extraction, and analysis strategy, we extracted the most relevant information in the proposed models. Our analysis revealed valuable findings about the field of research, such as the importance of waste data collection and positive influence on the prediction of waste generation after deploying prominent ML approaches such as K-nearest neighbors and support vector machines. Furthermore, the implications and limitations of our survey were highlighted to assist the keen researchers transparently.

Most of the works surveyed have targeted individual phases in the solid waste management life cycle. None of the studies developed an end-to-end waste management system with coordinated operations. Therefore, our future research activities include designing and implementing a comprehensive end-to-end smart waste management solution with the aim of covering all waste management processes and optimizing existing machine learning approaches, such as convolutional neural networks. Furthermore, collaboration with municipal authorities is highly commendable for future works, from the design phase to the implementation and testing phases, to realize a complete solution for smart waste management. Upcoming surveys should scrutinize the works modeling commercial waste, which is another important research area.