Evaluating Recycling Initiatives for Landfill Diversion in Developing Economies Using Integrated Machine Learning Techniques

Abstract

This study investigates the effectiveness of the Lagos Recycle Initiative (LRI) on landfill diversion (LFD) in Lagos, Nigeria, where evidence-based assessments of such initiatives are lacking. It evaluates the recycling diversion rate (RDR) of household recyclables (HSRs) across local government areas using field surveys and population data. Machine learning algorithms (logistic regression, random forest, XGBoost, and CatBoost) refined with Bayesian optimisation were employed to predict household recycling motivation. The findings reveal a low RDR of 0.37%, indicating that only approximately 2.47% (31,554.25 metric tonnes) of recyclables are recovered annually compared to a targeted 50% (638,750 metric tonnes). The optimised CatBoost model (accuracy and F1 score of 0.79) identified collection time and the absence of overflowing HSR bins as key motivators for household recycling via the SHapley Additive exPlanations (SHAP) framework. This study concludes that current LRI efforts are insufficient to meet recycling targets. It recommends expanding recovery efforts and addressing operational challenges faced by registered recyclers to improve recycling outcomes. The policy implications of this study suggest the need for stricter enforcement of recycling regulations, coupled with targeted financial incentives for both recyclers and households to boost recycling participation, thereby enhancing the overall effectiveness of waste diversion efforts under the LRI. This research provides a benchmark for assessing urban recycling initiatives (RIs) in rapidly growing African cities.

1. Introduction

Waste is generally defined as any material for which the generator has no further use either for the purpose of production, transformation, or consumption and is therefore discarded [1,2]. According to the recent report of the United Nations Environment Programme (UNEP), global municipal solid waste (MSW) generation is projected to increase from ~2.1 billion tonnes in 2020 to ~3.8 billion tonnes in 2050. The report shows Sub-Saharan Africa (SSA) as the region with the highest proportion of uncontrolled disposal (open dumping and open burning) of MSW in the global waste stream [3,4]. The implication of this is that, as industrialisation and urbanisation rates increase in SSA, the upsurge in MSW volume will cause a rise in the global share of uncontrolled disposal from the region [3]. In many developing countries (DCs) (countries with low levels of economic growth and inadequate technical and social infrastructures [5] like those in the SSA region), the uncontrolled disposal of MSW is primarily in unsanitary landfills without the segregation of recyclables from non-recyclables [6,7,8,9]. The disposal of co-mingled MSW in unsanitary landfills poses a threat to human health and the environment [10]. Moreover, the constraint of land for the construction of new landfill facilities in the urban areas of DCs is becoming more pronounced even as the current landfills are reaching their full capacity, which is further exacerbating the poor management of MSW in SSA and particularly in Nigeria [11].

Nigeria’s population was estimated at ~220 million as of 2022 [12], and is projected to be the third largest country by population globally, surpassing the population of the United States before 2050 [13]. Annual waste generation in Nigeria is ~30 million tonnes, with a paltry collection rate of 20–30% [14]. This implies that about 70% of the national waste stream containing valuable recyclables that could be diverted from landfills is lost through uncontrolled waste disposal. The continued rise in Nigeria’s population, among other factors, will cause an exponential rise in the volume of MSW in the coming years. In many Nigerian cities, including Lagos, MSW is disposed of primarily in landfills, which are mostly open dumps [15,16,17]. As of 2020, the population of Lagos, Nigeria’s largest city by population, was estimated at ~26 million [18] at a growth rate of ~3.2% higher than the national average of ~2.6% [19]. Lagos residents generate ~14,000 metric tonnes of waste daily [20], out of which ~3500 metric tonnes (25% of the waste stream) are recyclables like plastics, bottles, and paper [21]. At the current population growth rate, the waste management (WM) authority in Lagos would need to gear up efforts to match the pace of MSW collection with the rate of generation, especially prioritising the recovery and diversion of household recyclables (HSR) from its limited number of landfills. The Lagos State Government (LASG) is aware of the current challenges with landfill disposal and has rebranded its previous Blue Box Initiative (BBI) into the Lagos Recycle Initiative (LRI) in December 2020. The aim of the LRI is to encourage Lagos residents to recycle their waste as a way of contributing to the circular economy (CE) [22]. Using recycling as a tool to mop up recyclables generated through human activities contributes to minimising the impacts of mismanaged waste on the environment [23].

However, when compared to South Africa, Canada, and Australia where national waste diversion rates (WDRs) have been reported as 11% [24], 26% [25], and 64% [26], respectively, there is a paucity of data on Nigeria’s WDR, which could have been referenced in this study. Moreover, unlike in China where the national recycling rate is known to be below 20% (National Bureau of Statistics of China, 2023 as cited in [27]), there is no known record of recycling rate in Nigeria. To the best of the authors’ knowledge, no studies from any Nigerian city have ever evaluated the effectiveness of recycling initiatives (RIs) in Nigeria. This lack of data is not merely a statistical oversight but represents a significant gap in understanding the effectiveness of RIs in Africa’s most populous black nation. The dearth of information is of particular concern, given Nigeria’s rapid urbanisation and constant proliferation of MSW generation. Unlike its global counterparts, Nigeria, and indeed most of the Nigerian cities, have not been the subject of rigorous studies in evaluating the effectiveness of recycling programmes. This absence of empirical data hampers evidence-based policymaking and impedes the development of targeted, effective waste management strategies.

Therefore, this study aims to gather evidence on the performance of RIs in Nigeria regarding landfill diversion (LFD) and to identify the significant factors limiting their effectiveness in a bid to propose data-driven solutions. LFD (also known as waste diversion) is the process of reducing the amount of waste sent to landfills through reuse, recycling, composting, and other forms of recovery such as waste-to-energy [28]. It is hypothesised that the current LRI, despite its adoption, has not achieved substantial landfill diversion rates due to operational challenges and insufficient household participation. To test this hypothesis and effectively evaluate the LRI’s impact, this research has three specific objectives: (1) to determine the recycling diversion rate (RDR) of HSR in Lagos over a one-year period (January–December 2021); (2) to develop machine learning algorithms that can predict household motivation for recycling participation; and (3) to identify key factors influencing household recycling decisions using the SHapley Additive exPlanations (SHAP) framework. The novelty of this research lies in two key aspects. Firstly, the use of machine learning (ML) algorithms to evaluate the drivers of recycling motivation in Nigeria has not been explored. Secondly, there is a lack of empirical evidence on the performance of a city-wide recycling initiative in Lagos or any other Nigerian city using the RDR as an indicator. By focusing on Lagos, one of the economic hubs in Nigeria and Africa’s largest megacity in terms of population, this study provides a benchmark for assessing urban recycling initiatives in the context of rapidly growing African cities.

2. Literature Review

2.1. Recycling in Africa

The environmental impact of landfill disposal has become a critical concern worldwide, particularly in developing regions. RIs play a vital role in diverting waste from landfills, which not only conserves natural resources but also helps mitigate serious environmental risks associated with landfill leachate. Landfill leachate contains four major categories of pollutants: dissolved organic matter, inorganic macro-components, heavy metals, and xenobiotic organic compounds [29,30]. This toxic mixture poses significant threats to groundwater and surface water quality, with ammonia often persisting as a major long-term pollutant even as organic components decrease over time. The environmental hazards posed by landfills are particularly acute in developing nations where approximately 93% of waste ends up in open dumpsites, compared to just 2% in high-income countries [31]. This stark contrast underscores the urgent need for effective recycling programs (RPs) in African nations to reduce landfill dependency.

Several studies on recycling in Africa have focused on waste characterisation and composition analysis, informal waste sector contributions, policy challenges, and opportunities in recycling, as well as public perception and participation [32,33,34,35,36,37]. However, there is a critical gap in empirical assessments that quantify the impact of city-wide RIs on LFD on the continent, compared to the situation in the Global North. A few studies in South Africa [2,38,39] have also conducted an analysis of people’s perception of recycling and the determinants of their involvement. While the authors acknowledged the potential of achieving a higher LFD rate through increased investments in recycling infrastructure and more public engagement, as many other studies do, none of the studies presented empirical evidence of the RDR in the areas under consideration in this study. Similarly, a study in Kinshasha, Democratic Republic of Congo [40], analysed seven scenarios to determine the most optimised WM model for the city. After applying the lifecycle assessment approach (LCA), scenario 7, which adopted the process of waste valorisation through various treatment processes proved effective at cutting greenhouse gas emissions by half compared to scenario 4. Additionally, scenario 7 showed the potential to increase LFD rate up to 70%, although at a much higher cost to the city government. This result assumes that waste collection coverage in the city is at the 100% mark and confirms that efforts targeted at optimising the collection system can potentially increase the LFD rate. However, while an assumption like this provides insights for municipal authorities to improve their WM system, they do not replace the need for empirical evidence aimed at determining the actual RDR in a city. Furthermore, Kinobe et al. [41] examined the potential to recover recyclables from the city landfill in Kampala, Uganda using the reverse logistics system (reverse logistics refers to the pick-up of recyclables from the waste stream by traders who sell them directly to recycling industries). The authors quantified the various fractions of the waste stream at the site and found that a high proportion (about 63%) of recyclables are left in the landfill due to reasons such as limited materials processing power and the absence of a market for the recyclables, among others. This situation revealed existing WM challenges in DCs that prevented effective LFD through recycling. While it is in line with best recycling practice to recover recyclables at the source to achieve higher LFD rates and mitigate greenhouse gas emissions [42], the situation in Kampala further highlights the lingering challenges with MSW collection in Africa that persistently prevent a data-driven approach to determining the RDR of a recycling initiative. Furthermore, Louziz et al. [43], in their assessment of the WM challenges in Morocco, noted that to boost the government’s ambition of decreasing waste disposed in landfills, emphasis should be placed on enlightening the citizens on waste sorting at the source to achieve higher LFD. While the study added that collection at the source could potentially reduce landfill disposal by as much as 80%, a realistic analysis of what is diverted under an existing RI can provide reliable data on the effectiveness of such an initiative.

The success of RIs in enhancing LFD rates ultimately depends on infrastructure design that promotes sustainability and social well-being. Gardoni and Murphy [44] emphasised that infrastructure plays a critical role in ensuring community sustainability and resilience, as it directly influences individuals’ genuine opportunities to achieve different dimensions of well-being. They proposed a “society-based design” approach that considered not only the technical aspects of infrastructure but also how it affects societal outcomes, particularly for vulnerable populations who often face disproportionate impacts from environmental hazards. Applied to recycling infrastructure in African contexts, this perspective suggests that effective recycling systems must be designed with consideration of social justice elements, ensuring that benefits are equitably distributed and that recovery processes after disruptions provide opportunities for all community members. This holistic approach to recycling effectiveness connects the technical aspects of waste diversion with broader societal goals of sustainability and equitable development.

Recycling Effectiveness

The effectiveness of an RI or RP can be quantified using the recycling rate (RR), which can also take the form of WDR depending on how a country defines it [25,45,46]. WDR is the amount of waste diverted from disposal in landfills divided by the total amount of MSW generated, expressed as a percentage [45]. Waste diversion from landfills can be achieved through recycling, composting, or reuse [47], including waste-to-energy (WTE) [26]. The RR, also expressed as a percentage, is the amount of waste materials recycled from the total recyclables collected for diversion [47]. WDR is calculated as follows:

$$\mathrm{W}\mathrm{D}\mathrm{R}={\displaystyle \frac{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{W}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{e} \mathrm{D}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{e}\mathrm{d}(\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{y}\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{g}+\mathrm{r}\mathrm{e}\mathrm{u}\mathrm{s}\mathrm{e}+\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}+\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r} \mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{s} \mathrm{o}\mathrm{f} \mathrm{r}\mathrm{e}\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{y})}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{w}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{e} \mathrm{G}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d}}}\times 100\%$$

While WDR is a holistic measure of LFD, as shown in the above formula, it is possible to specifically assess only the recycling component, as shown in a study conducted in New York City (NYC) [48]. In that case, the study addressed the RDR, which is the proportion of the MSW stream that was collected for the purpose of recycling, expressed as a percentage [48]. In essence, the RDR is a component of the overall WDR. To incorporate this into the above formula, this is expressed as follows:

$$\mathrm{R}\mathrm{D}\mathrm{R}={\displaystyle \frac{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{w}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{e} \mathrm{d}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{f}\mathrm{o}\mathrm{r} \mathrm{r}\mathrm{e}\mathrm{c}\mathrm{y}\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{g}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{w}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{e} \mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d}}}\times 100\mathrm{\%}$$

In the NYC study [48], a city-wide assessment of 59 community districts was carried out by computing the tonnage per day of recyclables diverted for recycling in each district over the total waste generated. The resulting figure was then expressed as a percentage to determine the RDR for each district. In the current research, the LRI does not include the capture of materials diverted through composting, reuse, or WTE. Hence, the calculation is focused on the recycling component in the formula to determine the RDR, which aligns with the methodology used in the NYC study. A few studies have carried out a percentage quantification of HSR diverted from landfills. For example, Ali et al. [49] quantified fractions of MSW in Peshawar city in Pakistan, South Asia and obtained the value of 13.71% as the city’s WDR. The authors found that whereas the amount of recyclables, reusable materials, and organic components in the waste stream with a high potential for diversion was ~70%, ~86.8% were disposed of in landfills without recovery [49]. Like the work of Ali et al. [49], in Kumasi, Ghana, West Africa, Wahabu et al. [50] equally quantified the fractions of MSW diverted from landfills. In their study, they examined the impact of the Pay-As-You-Throw (PAYT) economic instrument on the amount of waste diverted from landfills through communal collection points where informal waste pickers are operating. Their findings showed that with the implementation of the PAYT, the deposit of HSR at the collection points was reduced. It was revealed that for each waste picker studied at the seven communal collection points, there was a contribution of ~6 metric tonnes of diversion rate of recyclables from landfills. When compared to the findings from the Pakistani study, it was observed that ~70% of the recyclables in the waste stream had the potential to be recovered and diverted from landfills. The Ghanaian study further noted that if the efforts of the waste pickers are aggregated across one hundred collection points, ~300 tonnes of recyclables could be diverted from landfills within a year [50].

Another study from London [46], United Kingdom, reported on home composting, an organic recycling process. In that study, a quantitative assessment of survey data obtained from 64 homeowners involved in home composting showed 20% WDR of biodegradable waste from landfills. Although the London study was carried out on a small scale, nonetheless, it contributed to the discourse on LFD, which can be applicable to many countries in SSA where the organic component of the MSW stream is the highest compared to other MSW fractions [51,52]. In another research study carried out in Brazil, South America, before the introduction of a private RI in a residential complex, all waste was landfilled with a zero per cent WDR. However, with the application of reverse logistics, coupled with effective implementation and monitoring over ten years, 67% of the total waste volume in the complex was diverted from landfill. This was achieved at no cost to the municipal government, nearly meeting the targeted 70% diversion rate [53]. The reverse logistics approach used in the Brazilian study is like the situation in many SSA cities, including Lagos, where the trade in recyclables is common.

In two Australian states (New South Wales and Victoria), a survey of some residential projects in the construction and demolition (C & D) sector showed a WDR of ~95% in each project compared to the Australian WDR average of ~64% [26]. However, while the study emphasised the significance of using WDR in assessing WM performance, the authors noted that the high WDR recorded in the study could not be guaranteed due to the non-reliability of waste data used for the analysis [26]. A study carried out by Pan et al. [25] in Canada, applied three novel metrics, which included WDR to assess the effectiveness of non-hazardous waste diversion models of public and private WM services in four Canadian provinces. The results of a time-series analysis used in the study showed that British Columbia had the highest average diversion rate of 32.7%. In their study, the authors suggested that WDR could be combined with other indicators in determining the effectiveness of an RI [25]. It is pertinent to note from these studies that in SSA, aside from the Ghanaian study where WDR among informal waste pickers was quantified, there is a lack of research on the effectiveness of city-wide RIs in SSA region, including Nigeria; hence, this study attempts to fill the research gap.

Beyond the environmental benefits of LFD, RIs can also create economic value through resource recovery. Yadav et al. [54] demonstrated that various types of waste—agricultural, industrial, and domestic—can yield valuable minerals and materials when properly processed. For instance, agricultural waste like rice husks and sugarcane bagasse can be sources of silica, while incense stick ash and eggshell waste can yield calcium-related products. In the context of SSA countries, where organic waste constitutes the highest fraction of MSW compared to other components [49], there are significant opportunities for value creation through composting and biofuel production. Such value-added approaches to waste management align with CE principles and could enhance the economic viability of recycling initiatives in African urban centres.

ML has recently started gaining traction for waste management applications globally. The implementation of smart city technologies offers promising solutions to waste management challenges in Africa. Kontokosta et al. [55] demonstrated that high spatial resolution waste generation estimates can significantly improve collection efficiency in urban environments. Their research developed an ML and small area estimation approach to predict waste generation at the building level in NYC, allowing more efficient collection routing and better targeting of behaviour change initiatives. This approach is particularly relevant for African cities facing rapid urbanisation and increasing waste volumes.

Chen [56] further explored how ML techniques combined with Internet of Things (IoT) technologies can enhance waste recycling processes in smart cities by automatically classifying and separating materials. The proposed Automatic ML-Based Waste Recycling Framework (AMLWRF) demonstrated the potential for IoT-powered devices installed in waste containers to provide real-time data on waste generation behaviour, achieving significant improvements in accuracy (96.1%), efficiency (97.1%), and recycling rates (91.6%) compared to conventional methods.

Building on these approaches, Erkinay Ozdemir et al. [57] comprehensively reviewed different ML algorithms being applied to recycling systems. They found that various algorithms including convolutional neural networks (CNNs), support vector machines (SVMs), decision trees, k-Nearest Neighbour (k-NN), and artificial neural networks (ANNs) have been successfully implemented in waste sorting and classification with accuracies often exceeding 93%. These algorithms have proven particularly effective in the separation of recyclable materials such as paper, plastic, metal, and glass, which is a critical step in improving recycling rates. As the authors noted, “The recycling rates are continuously increasing; however, assessments showed that humans will be creating more waste than ever before” [57], highlighting the urgent need for more efficient recycling systems powered by ML.

Despite the demonstrated effectiveness of ML applications in WM in developed countries—from predictive waste generation modelling to automated sorting and recycling—these technologies have rarely been applied in the African context, particularly for predicting household motivation for recycling. This represents a significant gap, as understanding the factors that drive recycling behaviour at the household level could inform more effective policies and interventions to increase recycling rates across Africa’s rapidly growing urban centres.

The review of these studies highlights a significant research gap in Africa, particularly regarding the evaluation of city-wide RIs. While the Ghanaian study by Wahabu et al. [50] quantified WDRs among informal waste pickers, there is a notable lack of comprehensive research on the effectiveness of large-scale RIs in African countries, including Nigeria. This scarcity of data and analysis presents a critical challenge in understanding and improving WM practices across the region. The present study aims to address this research gap by providing a detailed assessment of the LRI, a city-wide programme in Nigeria’s largest metropolis. Analysis of the extant literature on waste recycling in Africa further strengthens our position that there is currently a lack of empirical evidence on the evaluation of recycling initiatives on the continent. This study, rather than focusing on general WM discussions, closes this gap by providing a data-driven evaluation of the LRI, estimating the RDR and analysing the key factors for recycling motivation using ML techniques.

3. The Lagos Recycle Initiative

There had been other RIs in Lagos before the launch of the LRI. For example, in 2017, when the Environmental Management Protection Law of Lagos was signed into law, the Clean Lagos Initiative (CLI) was introduced by a former governor of Lagos State to achieve a more sustainable and cleaner Lagos [58]. Although the CLI had its own gains and challenges, it was short-lived and was replaced by the BBI, which was launched by the incumbent governor in 2019. The focus of the BBI was to ensure the capture of 50% of recyclables in the waste stream by 2021 through the source separation of domestic waste using two colour-coded bags: a blue bag for recyclables and a black bag for non-recyclable waste [20]. It is interesting to note that while the BBI was yet to become fully operational, it was rebranded into the LRI. The LRI aims to address the challenges of plastic pollution through recycling and the use of a financial incentive system to encourage participation. While the use of a financial tool to spur motivation for recycling has been widely used, a study in China showed that its effectiveness remains unclear [59]. During the launch of the initiative, the officials of the Lagos State Waste Management Authority (LAWMA) noted that Lagos State has limited landfill space, hence the need to proactively step up the recovery efforts on the collection of recyclables from the waste stream to reduce the amount of waste going to the landfills [60]. The initiative started as a pilot scheme in some organised housing estates in Lagos, where recyclables (mainly plastics) were collected in receptacles provided by LAWMA and sorted for sale to off-takers who processed them into other products [60]. The locations selected for the pilot scheme were the Lagos State Development and Property Corporation (LSDPC) Estates at Adeola Odeku, Victoria Island; Glover Street, Ebute Meta, and Akerele, Surulere. The rationale for the choice of these locations could probably be due to the organised nature of estate developments, which made such a scheme easier to test. As at the beginning of the second quarter of 2021, the programme had been extended to 42 Local Council Development Areas (LCDAs) out of the 57 LCDAs in the state with 48 recyclers who carried out the door-to-door collection of recyclables [22] (Figure 1). Although LAWMA stated that there are plans to extend the initiative to other LCDAs, we could not confirm how many more had been added at the time the field survey was carried out.

In 2021, the LAWMA commissioned ~400 aggregators at Olusosun landfill (the city’s largest landfill site), to ensure the proper aggregation of recyclables [21]. It is imperative to note that according to the LAWMA website, collectors of recyclables (registered recyclers) also function as aggregators. This information was further confirmed via a phone call with one of the registered recyclers.

The Pakam App

The Pakam app (PA) is an innovative approach to household recycling in Lagos. It connects all the stakeholders working on the LRI together—the households, recyclers (aggregators), and LAWMA [21]. According to LAWMA’s video explanation via its Instagram handle [62], to use the PA as a collector, an individual requires authorisation from a registered recycling company, after which an identity card would be issued. As soon as this procedure is completed, available recyclables in households can be picked up via a “pick up task” button in the app. Depending on the collector’s choice, specific tasks or all tasks can be accepted. An aggregator (or collector) is deemed to have completed the pickup of HSR once the recyclables are weighed in kilograms, and the data are transmitted to the LAWMA waste database via the PA. Households who do not use the PA to place their collection requests have the opportunity to bring their sorted HSR to community recycling centres (CRCs) operated by aggregators for weighing and payment. This ensures wider coverage for the collection of recyclables, especially for those who are not digitally savvy. Since the app is not utilised in this case, the aggregators collate the total number of collected recyclables at their CRCs and send the data manually to LAWMA after weighing. The app contains a wallet through which households can request payment from collectors for the recyclables collected from them by filling in their bank details. The amount due for payment is easily determined as the app automatically calculates the value of the recyclables. In situations where a recycler cannot meet up with a collection request due to a busy schedule, LAWMA has a backup collection plan (Figure S1) in all the 20 local government areas (LGAs), and this is to ensure that no HSR available for pick up is left uncollected. This information was confirmed during an interview with LAWMA’s Central Head of Recycling. All collated recyclables are then sold to off-takers for processing in recycling facilities.

4. Results and Discussion

4.1. Analysis of Data from the Recyclers

In this section, the results of two datasets are presented—one from the recyclers in Section 4.1 and the other from the households in Section 4.2. This is to ensure a clearer presentation of findings from the two stakeholders, given that two sets of questionnaire surveys were administered.

4.1.1. Distribution of the Collectors

Figure S2 illustrates the distribution of collectors of HSR across the studied areas. The data revealed that Surulere and Agege have the highest number of collectors, with three collectors each, while Eti-Osa has the lowest representation, with one collector. The equal distribution between Surulere and Agege may indicate a balanced approach to recycling implementation in these areas, which represent a mix of residential and commercial zones. This parity might reflect similar waste generation patterns or equally effective local implementation strategies in these LGAs. As noted earlier, the collectors surveyed in the three LGAs were names given to us by the Central Head of Recycling of LAWMA. Inappropriately, unlike Agege and Surulere, only one recycler appeared to be active in Eti-Osa LGA as of the time of the field survey. This could be due to Eti-Osa’s status as a high-income area with a predominantly busy working-class population who may not have the time or motivation to sort recyclables. Unlike areas such as Agege and Surulere, where there are more recyclers because there are probably more people interested in trading in recyclables, Eti-Osa has shown relatively low engagement in this practice.

4.1.2. Waste Bin Distribution Coverage

Figure S3 demonstrates the waste bin distribution coverage for recyclable waste storage across the studied areas. The majority of collectors (57.1%) reported a coverage between ~50% and 79%, indicating a substantial progress in establishing recycling infrastructure at the household level. However, 28.6% of collectors reported no involvement in bin distribution, suggesting potential gaps in implementation or areas where distribution was deemed unnecessary. A smaller portion (14.3%) reported coverage below 49%, highlighting areas that might require additional attention. Markedly, no collectors reported coverage of 80% or above, indicating that there was room for improvement even in well-served areas. These findings underscore both the achievements and challenges of the LRI’s bin distribution efforts, providing insights for targeted improvements to enhance recycling participation and LFD rates.

4.1.3. Waste Collection Frequency

The frequency of waste collection in the studied areas revealed significant patterns, as shown in Figure S4. A large majority of waste collectors (71.4%) conducted daily collections, indicating a high frequency of service. This may reflect the substantial volume of recyclable waste generated in these urban areas. In contrast, a smaller portion (14.3%) of collectors reported weekly collections, which might be attributed to lower waste generation in certain areas or logistical challenges that hinder daily collection. Additionally, 14.3% did not provide data on collection frequency, potentially indicating inconsistent collection schedules. The dominance of daily waste collections suggested that the initiative aims to maintain a consistent and frequent waste removal service, a critical factor in encouraging continued participation in recycling and managing urban waste effectively. Nevertheless, the observed variation in collection frequency underscores the need for adaptable WM strategies to address the specific needs of different areas in Lagos.

4.1.4. Truck Capacity and the Weekly Volume of HSR Collected

Figure S5 presented the comparison between total truck capacity and the weekly volume of HSR waste collected by each collector. The data revealed noticeable variations among collectors in both their truck capacities and collection volumes. Collector 2 stands out with the highest truck capacity (75 metric tonnes) and weekly collection volume (62.5 metric tonnes), indicating efficient resource utilisation. Most collectors demonstrated underutilisation of their truck capacities, with weekly collection volumes significantly lower than their potential capacity. This disparity is particularly notable for collectors 6 and 7, who had substantial truck capacities (60 and 15 metric tonnes, respectively) but collected only 2.0 and 0.5 metric tonnes weekly. Interestingly, collector 5 did not provide a response regarding truck capacity during data collection but reported a weekly collection volume of 0.5 metric tonnes, suggesting possible use of alternative collection methods or data reporting inconsistencies. These findings highlighted opportunities for optimizing resource allocation and improving collection efficiency across the initiative, potentially through better route planning, increased community engagement, or reallocation of resources to match actual collection needs in different areas.

4.1.5. HSR Material Type Collected

The distribution of collectors by HSR material type revealed distinct patterns in the material collection (Figure S6). Plastics were the most frequently collected material, with all collectors (100%) participating in their collection. Used beverage containers (UBCs) followed, with five collectors (71.4%) engaged in their collection, while four collectors (57.1%) handled paper. Glass was collected by only three collectors (42.9%), indicating it was less commonly targeted. The prominence of plastics in collection efforts highlights their significance in the urban waste stream and their central role in the recycling economy. The variations in collection across other materials likely reflected differences in material prevalence, recyclability, market demand, or processing capacity. These findings suggests potential opportunities to expand and diversify recycling efforts, capturing a wider range of materials and enhancing the effectiveness of the LRI.

4.1.6. HSR Collected Versus Targeted

Figure 2 illustrates the actual versus target monthly HSR waste collection for each collector in the LRI, measured in metric tonnes. The data showed significant variations in both collection targets and performance across collectors. Collector 2 demonstrated optimal performance, achieving 100% of their substantial 250-metric tonnes target. Similarly, collectors 3 and 6 also reached 100% of their targets, collecting 10 and 8 metric tonnes, respectively. However, other collectors showed varying degrees of underperformance. Collector 4 achieved 30% of their target, collecting 6 out of 20 metric tonnes, while collector 7 met 20% of their goal, gathering 2 out of 10 metric tonnes. The most notable discrepancies were observed with collectors 1 and 5; collector 1 achieved only 0.64% of their target, collecting 0.96 metric tonnes against a substantial 150-metric tonnes goal, while collector 5 reached 6.67% of their target, gathering 2 out of 30 metric tonnes. These disparities highlighted the challenges in setting and achieving collection targets across different areas or operational contexts within Lagos. The data suggests a need for reviewing and potentially adjusting targets, as well as investigating and addressing factors contributing to underperformance in some areas to enhance the overall effectiveness of the RP.

4.1.7. Challenges Faced by the Collectors

The distribution of challenges encountered by collectors in the LRI revealed several critical issues, as depicted in Figure S7. Logistics is the most common challenge, affecting 57.1% of collectors, and reflecting significant operational difficulties in waste collection and transportation. Additionally, two challenges—lack of awareness and extortion by hoodlums—were reported by 28.6% of collectors, indicating substantial barriers related to community engagement and security. Other challenges affecting 14.3% of collectors, include a range of issues such as insufficient funds, inadequate equipment, competition, corruption, multiple taxation, inability to meet set-targets, inadequate storage, non-inclusivity, and lack of truck ownership. This wide variety of challenges highlights the complexity of implementing an efficient RP. While logistics appeared to be a common issue, many other challenges might be more localised or specific to individual collectors. The diverse nature of the challenges underscores the need for a comprehensive, multi-faceted approach. Hence, to address the underperformance of recyclers under the LRI, policymakers could provide financial support, leverage data-driven approaches to optimise collection logistics, and invest in the waste-sorting infrastructure to improve the operational efficiency of the recyclers.

4.1.8. Estimation of the RDR

In estimating the RDR from the formula stated earlier, the total annual tonnage of HSR collected from the three LGAs was divided by the estimated total waste (annual) generated in the LGAs in 2021 (the assessment year). The total tonnage (annual) was obtained by converting the weekly tonnage collected by each recycler to an annual figure (

Table 1. The tonnage of recyclables collected by recyclers in the three LGAs.

Respondent’s ID	Weekly Tonnage (Metric Tonnes)	Estimated Annual Tonnage (Metric Tonnes)
Collector 1	0.24	12.48
Collector 2	62.5	3250
Collector 3	2.5	130
Collector 4	1.5	78
Collector 5	0.5	26
Collector 6	2	104
Collector 7	0.5	26

). This compares well with the methodology used in the NYC study [48] where an average daily tonnage (ADT) of HSR and ADT of the total waste generated in each of the 59 districts was cumulated to annual figures to arrive at an estimate of the RDR for each district. The data used in the NYC study were obtained directly from the city’s sanitation department. In the current research, since the data on the total waste generated in the LGAs in 2021 were not available, the 2019 population figures from the Lagos Bureau of Statistics (LSBS) [63] were extrapolated to estimate the 2021 population for each LGA using the geometric growth model formula [P_t = P_o (1 + r)^t] [64]: where P_t is the projected population at time, t; P_o is the initial population; r is the annual growth rate expressed as a decimal; and t is the number of years into the future.

This method is commonly used in demographic and WM studies for estimating population over short periods [64,65,66]. The total daily waste generated in each LGA in 2021 was then estimated by multiplying the 2021 population of the LGAs with the waste per capita/day for Lagos (0.51 kg/capita per day), which is the average for cities in SSA [3,67]. The resultant data were converted to annual values and summed up for the RDR calculation. The estimated RDR from the three LGAs was 0.37% (this represents the proportion of HSR in the annual MSW waste stream in Lagos that was diverted from landfills in 2021). This implies that of the officially reported 3500 metric tonnes per day of recyclables (1,277,500 metric tonnes/year) in the waste stream, the three LGAs contributed to the recovery and potential diversion of ~4730 metric tonnes of HSR from landfills in one year. Assuming the data are taken as a yearly average for every three LGAs in Lagos, cumulatively, it means recyclers in the 20 LGAs could potentially recover about 31,554.25 (2.47%) metric tonnes of HSR from the estimated 1,277,500 metric tonnes/year of HSR in the waste stream. This assumption may not hold true for all the LGAs due to factors like the number of recyclers in each LGA and their operational efficiency, among other considerations. In any case, it showed that the LRI, if well monitored, can continue to contribute to increasing LFD rate for recyclables. However, comparing the 2.47% with LAWMA’s targeted recovery of 50% (638,750 metric tonnes/year) of recyclables from the waste stream by June 2021, this figure is far below the projection and calls for an urgent review of the city’s approach to the recovery of HSRs.

Undoubtedly, it is possible that sometimes targets set by municipal waste authorities might be confronted with challenges which impeded the achievement of the recycling targets. For instance, in the study carried out in NYC [48], the city sanitation department in New York (whose role was closely comparable to that of LAWMA in Lagos) had targeted the recovery of about 50% recyclables from the city’s daily waste generation of 13,000 metric tonnes [68]. Due to the sudden changes made to the city’s recycling programmes by the government, among other reasons; recycling participation dropped, and that negatively impacted the RDR bringing it below the previous rates. It is worth noting that the sanitation department did a lot regarding the education of the public on RIs. Also, it was ensured that recyclables were collected consistently weekly from the neighbourhoods within the city, all in a bid to strengthen recycling participation. Although NYC has a lower population and a more advanced WM system compared to Lagos, the waste generation per day and recycling target were related. While Lagos residents generated ~14,000 metric tonnes of MSW per day, LAWMA aimed to have achieved a recycling target of 50% since 2021, but this is way below the projection, given the abysmally low RDR from the three LGAs. Despite the challenges discussed in the NYC study, RDR values ranging from 9% to 31% were achieved in each of the 59 districts, while about 20% was realised for the entire NYC compared to 0.37% from the three LGAs in Lagos and the hypothetical 2.4% from all the LGAs. This indicated that the challenges could be surmounted. Aggregating the collection efforts of recyclers across the 20 LGAs in Lagos, like the projected combined efforts of waste pickers studied by Wahabu et al. [50], could contribute substantially to increasing the tonnage of recyclables collected in Lagos. To raise the RDR above current rates, as highlighted in the preceding section, the challenges faced by the recyclers need to be addressed. A good starting point could be LAWMA’s increased engagement with its registered recyclers to understand critical challenges limiting their collections below monthly targets and finding joint solutions.

4.2. Analysis of Data from the Households

4.2.1. Profile of the Respondents

The demographic profile of household respondents in the LRI is presented in Figure S8. Regarding the age distribution, the majority of respondents (58.0%) fall within the 30–49 age group, indicating a strong representation of working-age adults. This was followed by the 50–64 age group at 23.5%, suggesting significant participation from older adults as well. Younger adults (18–29) comprised 14.3% of respondents, while those at 65 and above represented 4.2%, showing a diverse age range in the study. Figure S8b revealed a highly educated sample, with 66.9% of respondents having attained tertiary education. Secondary school certificate holders made up 24.6% of the sample, while 8.5% had primary school certificates. The educational profile suggested that the initiative had engaged a predominantly well-educated segment of the population. The working status of respondents showed a diverse distribution. Full-time workers constituted the largest group at 38.8% of the sample, followed closely by those who owned businesses at 27.6%. The two categories accounted for 66.4% of respondents, representing a significant majority of employed and self-employed individuals. Part-time workers made up 12.9% of the sample, while those not working represented 14.7%. Job applicants constituted the smallest group at 6.0%. The distribution reflected a broad mix of employment statuses among participants in the RI, with a strong representation of full-time workers and business owners, balanced by a notable presence of part-time workers, unemployed individuals, and job seekers.

4.2.2. Data Preprocessing

In this study, a total of 137 questionnaires were distributed to the households of which 127 were considered valid for analysis. Many studies have also used similar data sizes to build efficient ML models in different domains [69,70,71].

Table 2. Encoding and description of the datasets.

Variables	Codes	Description	Type	Unique Values
Age	Age	Age of the respondent	Ordinal	18–29 (1)
				30–49 (2)
				50–64 (3)
				65 and above (4)
Working status	Wks	Employment status of respondent	Categorical	Full-time worker Part-time worker Not working Own business Job applicant
Educational level	Edl	Highest educational level of respondent	Ordinal	Pry. Schl. cert. (1) Sec. Schl. cert. (2) Tertiary inst. (3)
LRI awareness	Aws	Awareness of respondent about LRI	Binary	Yes (1) No (2)
Type of waste sorted	Tws	Types of HSR sorted by respondent	Categorical	Plastics Paper UBCs Glass
Frequency of HSR pick up	Fwp	Frequency of HSR collection	Ordinal	Once in 3 weeks (1) Once in two weeks (2)
				Once a month (3)
				Once a week (4)
				Twice a week (5)
No overflow of HSR bin	Nob	Respondent’s agreement with statement that HSR bin does not overflow	Ordinal	Strongly agree (5) Agree (4) Neutral (3) Disagree (2) Strongly disagree (1)
Regularity of HSR collection	Rwc	Whether HSR collection is regular	Binary	Yes (1) No (2)
Adequacy of HSR bin received	Awb	Whether HSR bin received by respondent is adequate	Binary	Yes (1) No (2)
Service quality during HSR collection	Sqc	Rating of service quality during HSR collection	Ordinal	Excellent (3) Fair (2) Poor (1)
Communication quality during HSR collection	Cqc	Rating of communication quality during HSR collection	Ordinal	Excellent (3) Fair (2) Poor (1)
Usage of Pakam app	Upa	Whether respondents use Pakam app during HSR collection	Binary	Yes (1) No (2)
HSR weighing at collection point	Hwc	Whether collectors weigh HSR at collection point	Binary	Yes (1) No (2)
Perception of household about recycling	Phr	Whether household perceives recycling as beneficial	Binary	Yes (1) No (2)
Perception of recycling productivity	Prp	Whether household perceives recycling as productive/time wasting	Binary	Yes (1) No (2)
Societal influence on recycling	Sir	Whether societal views on importance of recycling influence respondent’s recycling behaviour	Binary	Yes (1) No (2)
HSR collection time	Hct	Whether the time of HSR collection is regular	Binary	Yes (1) No (2)
Motivation to recycle	Mtc	To assess if respondents are motivated to recycle or not	Binary	Yes (1) No (2)

presents a detailed overview of the variables used in the analysis, including their codes, descriptions, data types, and unique values. The data preprocessing involved careful encoding of variables and standardisation to ensure optimal analysis [72,73], as follows:

• Ordinal data: Variables such as age, education level (Edl), frequency of HSR pick up (Fwp), no overflow of HSR bin (Nob), service quality during HSR collection (Sqc), and communication quality during HSR collection (Cqc) were encoded using label encoding. This method assigned numerical values to ordered categories, preserving the inherent order of the data (Table 1).
• Binary data: Variables like LRI awareness (Aws), regularity of HSR collection (Rwc), adequacy of HSR bin received (Awb), usage of Pakam app (Upa), HSR weighing at collection point (Hwc), perception of household regarding recycling (Phr), perception of recycling productivity (Prp), societal influence on recycling (Sir), HSR collection time (Hct), and motivation to recycle (Mtc) were encoded as binary values (1 for Yes, 0 for No).
• Categorical data: For variables such as working status (Wks) and type of waste sorted (Tws), one-hot encoding was employed. This technique created separate binary columns for each category, allowing for non-ordinal categorical data to be effectively used in various analytical models.
• Data standardisation: After encoding, numerical features were standardised to ensure all variables contribute equally to the model and to improve the convergence of ML algorithms. The standardisation process involved scaling the features to have a mean of 0 and a standard deviation of 1, using Equation (1):

$$X_{standardized}={\displaystyle \frac{(X-\mu )}{\sigma }}$$

(1)

where $X$ is the original feature vector; $\mu$ is the mean of the feature; and $\sigma$ is its standard deviation.

4.2.3. Results of the ML Models

The performance of the base and optimised ML models on the testing dataset is presented in

Table 3. Results of the base models on the testing dataset.

Models	Accuracy	Weighted Precision	Weighted Recall	Weighted F1 Score
LR	0.67	0.70	0.67	0.68
RF	0.50	0.62	0.50	0.53
XGBoost	0.51	0.63	0.50	0.53
CatBoost	0.75	0.75	0.75	0.75

and

Table 4. Results of the optimised models on the testing dataset.

Models	Accuracy	Weighted Precision	Weighted Recall	Weighted F1 Score
LR + BO	0.75	0.75	0.75	0.75
RF + BO	0.75	0.75	0.75	0.75
XGBoost + BO	0.75	0.75	0.75	0.75
CatBoost + BO	0.79	0.78	0.79	0.79

, respectively. Among the base models, CatBoost demonstrated the highest performance across all metrics, achieving an accuracy, weighted precision, weighted recall, and weighted F1 score of 0.75. LR showed moderate performance with an accuracy of 0.67 and a weighted F1 score of 0.68. In contrast, RF and XGBoost models exhibited lower performance, with accuracies of 0.50 and 0.51, respectively, and a weighted F1 score of 0.53. These results suggested that, without optimisation, CatBoost was the most effective in capturing the underlying patterns in the household recycling behaviour data. The application of the BO algorithm led to substantial improvements in the model performance across all algorithms. Table 3 showed that the optimised models achieved consistently higher metrics compared to their base counterparts. Interestingly, LR + BO, RF + BO, and XGBoost + BO all converged to an identical performance level, with accuracy, weighted precision, weighted recall, and a weighted F1 score of 0.75, matching the performance of the base CatBoost model. This convergence suggested that the dataset might have inherent characteristics or limitations that cap the predictive power at that level for most models. It could indicate that the algorithms when optimised were capturing similar underlying patterns in the data, possibly reaching the maximum predictive capacity given the available features and data quality [74,75]. The optimised CatBoost model (CatBoost + BO), however, managed to surpass that apparent threshold, achieving the highest overall performance with an accuracy of 0.79 and a weighted F1 score of 0.79. These findings underscored the importance of hyperparameter optimisation in enhancing model performance and demonstrated that CatBoost, when optimised, provided the most accurate predictions for household participation in the LRI. The superior performance of CatBoost + BO might be attributed to its ability to handle categorical variables effectively and its robust handling of complex interactions within the data [75,76].

Figure S9 presented the precision–recall curves for the four optimised models. The precision–recall curve is a valuable tool for evaluating the performance of classification models, especially in scenarios with imbalanced datasets. It plotted the trade-off between precision (the proportion of true positive predictions among all positive predictions) and recall (the proportion of true positive predictions among all actual positive instances) at various classification thresholds. The average precision (AP) score, shown in parentheses for each model, represented the area under the precision–recall curve. Higher AP scores indicated better overall performance. In this case, LR + BO and CatBoost + BO showed the highest AP of 0.87, followed closely by XGBoost + BO (0.85) and RF + BO (0.81). The curves for all models started at high precision and low recall, gradually decreasing in precision as recall increases, which was the typical pattern for precision–recall curves. Comparing the models, CatBoost + BO (red line) maintains the highest precision across most recall values, indicating robust performance. LR + BO (blue line) showed a strong performance, particularly at lower recall values where it maintained very high precision. XGBoost + BO (green line) performed well, often matching or closely trailing the top performers, while RF + BO (orange line) generally showed lower precision compared to the other models, especially at higher recall values.

4.2.4. Factors Influencing Household Motivation to Recycle

The SHAP analysis of the best-performing model, CatBoost + BO, provided valuable insights into the factors that most significantly influence household motivation to participate in the LRI. Figure 3 and Figure 4 illustrated the importance of the feature and its impact on the model’s predictions for the top 20 features, respectively. The Hct emerged as the most influential factor by a significant margin, suggesting that the regularity of waste collection services plays a crucial role in motivating households to participate in recycling programmes. The timely collection of waste is essential in the WM process [77]. Where there is an irregularity in the collection, bins tend to overflow and can result in the inappropriate segregation of recyclables at the source [78]. When this happens repeatedly, households may become demotivated to recycle over time. Following Hct, the next most important features are the Nob, Upa, and Phr. These factors underscore the importance of proper WM infrastructure, technology adoption, and household attitudes toward recycling in driving participation behaviour. Regarding the essence of WM infrastructure, one study [79] observed that the availability of waste segregation facilities tend to improve waste sorting, thereby improving the separate collection of recyclables. The visualisations also revealed that Edl, Fwp, and age had moderate impacts on recycling participation. Interestingly, the employment status, particularly part-time work (Wks_Part time worker), showed a notable influence, ranking higher than some service-related factors. Sqc, Aws, and Cqc showed relatively minor influences yet still contributed to the model’s predictions. It is pertinent to note that while the two dominant factors of Nob and Hct may be peculiar to the situation in Lagos under the LRI, some scholars in SSA have identified other significant factors such as household income, formal education, access to social grants, and age as motivators of recycling participation [80]. This implies that under diverse conditions and circumstances, people’s motivations for recycling might differ. As such, it is essential to understand the local situation when making policies aimed at improving recycling outcomes.

The SHAP values plot in Figure 4 provided a more nuanced view of each feature’s impact, with the colour gradient from blue to red representing low to high feature values. HSR collection time displayed a clear pattern where longer collection times strongly decreased participation likelihood, while shorter times increased it. The absence of overflow in HSR bins showed a positive correlation with participation, suggesting that adequate WM capacity encouraged engagement. Remarkably, the usage of the Pakam app exhibited a negative correlation with SHAP values, indicating that higher app usage is associated with lower participation likelihood. This unexpected result might warrant further investigation into the effectiveness of the app or user experience. Combining the theory of planned behaviour (TPB) and the technology assessment model (TAM), a study on the use of mobile apps to encourage environmental behaviour like recycling observed that two predictors of app use intentions (attitude towards the app and its perceived usefulness) should be considered when promoting the usage of such apps [81]. While the study concluded that using a green app can encourage recycling behaviour, the negative correlation observed with the Upa, might necessitate the need to carry out a study assessing factors that can motivate increased Upa under the LRI. Such factors, coupled with awareness, are then targeted when promoting the app, with a view to increasing usage.

The perception of households regarding recycling showed a positive impact when values were high, reflecting the influence of favourable attitudes on recycling behaviour. This finding is consistent with the results of a study in one of the LGAs in Lagos, where it was shown that when people perceive recycling as beneficial to the environment, they tend to show a positive attitude towards it to avoid its negative impacts on the environment [82]. Edl demonstrated a mixed effect, with both positive and negative impacts across different education levels, suggesting a complex relationship between education and recycling participation among the sampled population. Evidence from some studies showed that, as the educational level rises, the interest in pro-environmental behaviour increases correspondingly [83,84]. The negative impacts observed here suggested that this might not always be the case, as a study [85] found that the percentage of educated individuals in a population did not necessarily influence the willingness to engage in pro-environmental behaviour.

The frequency of HSR pickup showed a slightly positive influence, emphasizing the importance of delivering quality WM services. Tabernero et al. [86] affirmed that a strong relationship exists between service quality and recycling behaviour. Their research showed that individuals who were satisfied with the services offered by a recycling provider would show a higher willingness to recycle compared to those who did not. It was then suggested that other studies could test the frequency of pickup as one of the objective indicators of service quality. The findings from the current research affirmed that the frequency of pickup was one of those measures employed by households to test the quality of service they received from providers of recycling services. Thus, recyclers in Lagos might need to engage more with their customers on grey areas of service improvement. Age also displayed a mixed effect, with both positive and negative impacts across different age groups, indicating that the relationship between age and recycling participation was not straightforward. Although, 72.3% of respondents in this study were aged 30–49 and only 27.9% were 50 or older, 78.9% of the older group demonstrated active involvement in recycling compared to the younger respondents. This is consistent with previous studies, which have found that older adults tend to exhibit greater interest in pro-environmental behaviours compared to younger individuals [83,86]. However, in future studies, the proportion of older people sampled could be increased to an appreciable level to better understand the relationship between age and recycling in the studied areas.

4.3. Sensitivity Analysis and Cross-Validation

Sensitivity analysis is crucial for evaluating model robustness and determining the optimal features for deployment in real-world applications. This section presents the results of feature subset sensitivity analysis, sample size sensitivity analysis, and cross-validation for the optimised CatBoost model that achieved the highest performance in predicting household recycling motivation in Lagos.

4.3.1. Feature Subset Sensitivity Analysis

Feature subset sensitivity analysis was conducted to examine how model performance varies when key features are included or excluded from the predictive model. The analysis focused on the five most important features identified through SHAP analysis: Hct, Nob, Upa, Phr, and Edl.

Table 5. Result of feature subset sensitivity analysis using the selected model.

Feature_Setup	Included_Features	Excluded_Features	Weighted_Accuracy	Weighted_Precision	Weighted_Recall	Weighted_F1_Score
All Top 5	Hct, Nob, Upa, Phr, Edl	None	0.791667	0.781579	0.791667	0.785012
Excluding Hct	Nob, Upa, Phr, Edl	Hct	0.75	0.5625	0.75	0.642857
Excluding Nob	Hct, Upa, Phr, Edl	Nob	0.75	0.725	0.75	0.731579
Excluding Upa	Hct, Nob, Phr, Edl	Upa	0.791667	0.781579	0.791667	0.785012
Excluding Phr	Hct, Nob, Upa, Edl	Phr	0.791667	0.781579	0.791667	0.785012
Excluding Edl	Hct, Nob, Upa, Phr	Edl	0.791667	0.781579	0.791667	0.785012

reveals significant insights into the relative importance of each feature for predicting household recycling motivation. When all five top features are included, the model achieves optimal performance with a weighted accuracy of 79.17% and an F1 score of 78.50%. The exclusion of Hct causes the most substantial performance degradation, with weighted precision dropping dramatically from 78.16% to 56.25% and F1 score declining from 78.50% to 64.29%. This confirms Hct as the most critical predictor of recycling motivation, consistent with its highest SHAP importance value.

Excluding the Nob feature also negatively impacts model performance, though less severely than Hct, with weighted precision dropping to 72.50% and F1 score to 73.16%. This indicates that adequate waste bin capacity is a secondary but still important predictor of recycling participation. Interestingly, the exclusion of Upa, Phr, or Edl individually has no discernible impact on model performance, with metrics remaining identical to the full model. This suggests a potential correlation among these features, where information lost from excluding one feature can be compensated by the remaining features [87]. These findings demonstrate that Hct and Nob are the most critical operational factors influencing recycling motivation among Lagos households. When waste is collected at regular intervals, and bins do not overflow, households are significantly more likely to participate in recycling programs. The negligible impact of excluding Upa aligns with the unexpected negative correlation observed in the SHAP analysis, suggesting that improving the Pakam app’s effectiveness or user experience might be necessary to enhance its contribution to recycling participation.

4.3.2. Sample Size Sensitivity Analysis

Sample size sensitivity analysis evaluates how model performance varies with different training/test split ratios. This analysis is essential for understanding the data volume necessary for reliable predictions and assessing model generalisability across different test set sizes.

The results in

Table 6. Result of sample size sensitivity analysis.

Train Size	Test Size	Weighted Accuracy	Weighted Precision	Weighted Recall	Weighted F1 Score
0.3	0.7	0.82	0.83	0.82	0.81
0.5	0.5	0.83	0.83	0.83	0.82
0.7	0.3	0.77	0.77	0.77	0.76
0.9	0.1	0.75	0.74	0.75	0.73

reveal an interesting pattern: model performance on test data generally improves as the test set size increases. With a small test set (10% of data), the model achieves 75% weighted accuracy and 73% F1 score. Performance increases as the test set grows, with strong results at both the 50/50 split (83% accuracy, 82% F1 score) and the 30/70 split (82% accuracy, 81% F1 score). This relationship between test set size and performance metrics merits careful consideration. First, larger test sets typically provide more reliable performance estimates by capturing a wider range of household characteristics and recycling behaviours present in the population. The improved performance with larger test sets suggests that these sets better represent the diversity of recycling motivations in Lagos households. Second, the smaller test sets (particularly the 10% configuration) may suffer from increased variance in performance metrics due to their limited size. With only approximately 13 samples in the smallest test set, individual misclassifications can disproportionately affect overall performance metrics, potentially yielding less stable assessments of model performance. Third, the relatively consistent performance between the 50/50 and 30/70 splits indicates that even with a reduced training set (30% of data, approximately 38 samples), the model can effectively learn the key relationships governing recycling motivation. This suggests that the primary factors influencing household recycling behaviour in Lagos may be relatively straightforward and can be captured without extensive training data. The optimal performance at the 50/50 split represents a balanced approach that provides both adequate training data for model learning and a sufficiently large test set for reliable performance evaluation. This finding has practical implications for future research on recycling behaviour in developing economies, suggesting that moderate-sized, carefully collected datasets may be sufficient for identifying key motivational factors. These results also highlight the importance of careful consideration of test set size when evaluating model performance, particularly in studies utilizing convenience sampling in complex urban environments where demographic and socioeconomic factors may vary considerably across locations.

4.3.3. Cross-Validation

To validate the robustness of the CatBoost model and ensure that performance metrics were not artifacts of a particular data split, a k-fold cross-validation (k = 5) was implemented. Cross-validation provides a more comprehensive assessment of model performance by evaluating the model on multiple subsets of the data, thus mitigating the risk of overfitting and yielding more reliable performance estimates [88]. Figure 5 presents the cross-validation performance of the optimised CatBoost model across four key metrics: weighted accuracy, weighted precision, weighted recall, and weighted F1 score. Each bar represents the mean performance metric over five folds, with error bars indicating the standard deviation of each metric across folds. The cross-validation results demonstrate strong and consistent performance across all metrics, with weighted precision showing the highest value at 0.859 ± 0.016. Weighted accuracy and weighted recall both achieved identical mean scores of 0.823 ± 0.027, indicating balanced performance between correctly identifying positive and negative instances of recycling motivation. The weighted F1 score, which represents the harmonic mean of precision and recall, showed a solid performance of 0.811 ± 0.030. Importantly, the relatively small standard deviations (ranging from 0.016 to 0.030) across all metrics indicate high stability in model performance across different data partitions. This consistency is particularly important given the convenience sampling method used in data collection, as it suggests that the model’s predictive capacity is not overly sensitive to the specific subset of households included in the training or testing sets. The cross-validation results confirm and strengthen the findings from the initial model evaluation. The optimised CatBoost model demonstrates reliable performance in predicting household recycling motivation, with weighted precision consistently exceeding 85%. This high precision is particularly valuable in the context of recycling program management, as it indicates that the model is especially effective at correctly identifying households likely to participate in recycling initiatives, allowing for more targeted interventions and resource allocation. The robust cross-validation performance further validates the importance of the key features identified through SHAP analysis. The stability of model performance across different data subsets reinforces the conclusion that Hct and Nob are indeed critical factors influencing household recycling motivation in Lagos, and not merely artefacts of a particular data partition.

These findings provide solid evidence that despite the limitations of the dataset size and sampling methodology, the identified factors influencing recycling motivation represent genuine patterns in household behaviour that can reliably inform policy and operational improvements for the LRI.

5. Methodology

5.1. Design

This study employed the quantitative research method (QRM) to evaluate the impact of the LRI on LFD by providing data-driven insights into the effectiveness of recycling in Lagos. A questionnaire survey was conducted among selected households and collectors of HSRs across three LGAs to gather quantitative data. At the same time, ML was applied to determine the significant factors of recycling motivation. The QRM involves the selection of a sample from a population using one or more sampling techniques; data obtained from the sample are then analysed and verified with appropriate statistical tools to make generalisations about the larger population from where the sample was taken. The choice of this method is predicated on the fact that the structured survey and ML models can be reproduced, implying that other researchers can either validate or refute the results due to its structured methodology [89]. One of the strengths of this approach is the systematic and objective collection of quantifiable data from the sampled population [90]. Also, the findings can be generalised across different contexts [89], thus providing insights to policymakers based on empirical evidence. However, a limitation of the QRM is sometimes in the improper representation of the target population when the collected sample is compared with the larger population, which then impedes the possibility of generalising the findings [91]. Some other weaknesses of the method are the high cost of gathering field data for large-scale studies and the limited contextual understanding, which results in overlooking essential nuances, thereby limiting the depth and interpretation of research findings [89].

5.2. The Study Area

Lagos is situated on an area of 351,861 hectares and has five administrative divisions (AD) (including Ikeja, Badagry, Ikorodu, Lagos (Eko), and Epe) [92]. The locations selected for this study are three of the twenty local government areas in Lagos (including Agege, Eti-Osa, and Surulere LGAs). Both Surulere and Eti-Osa LGAs fall under the Lagos (Eko) AD, while Agege is under the Ikeja AD [92]. Socio-economically, Eti-Osa LGAs are mostly household income earners, while Surulere and Agege LGAs can be categorised as medium and low-income areas, respectively [93]. According to the LSBS [94], both Agege and Surulere LGAs have higher population density compared to Eti-Osa. An examination of waste collection data in the three LGAs published in 2016 by the LSBS [94] showed that the volume of MSW collected in Eti-Osa is about 16 times the volume collected in Agege and Surulere LGAs as of 2015. A further investigation of the data indicated that Eti-Osa appeared to have more waste pick-up points compared to the other two locations. This could be attributed to the more organised nature of buildings in Eti-Osa, which makes accessibility easier for waste collection companies. The basis for the selection of the study areas is that two of the three LSDPC estates, where the pilot scheme for the LRI was tested, were in Surulere and Eti-Osa LGAs. Agege LGA was included because some active recyclers whose contacts were obtained from the LAWMA officials and were interviewed, operated in that area. Hence, conducting field surveys in locations where the LRI had been tested and where active recyclers are currently operational was considered very appropriate. Figure 6 shows the map of Nigeria and indicates the location of Lagos.

Institutional Framework for Waste Management in Lagos

The WM system in Lagos comprises a network of formal and informal institutions operating from the urban to the household level (Figure 7). At the urban level, the Lagos State Ministry of the Environment and Water Resources (MOE) is responsible for the overall management of the environment and all environmental-related matters in the State [96]. To ensure effective and efficient environmental administration, the MOE currently supervises 11 state-owned agencies, one of which is the LAWMA, the sole regulator of the WM sector. In LAWMA’s bid to bring its services to the grassroots, it concessions public and domestic WM operations, landfill operations, and transfer loading station responsibilities to private sector partnerships (PSPs), while it guides, models, and appraises the performance of these practitioners in line with international best practices [97]. The PSPs carry out the door-to-door collection of waste (mainly general waste) across the 20 LGAs, while the registered recyclers focus on the collection of recyclables in areas assigned to them. Both the PSPs and registered recyclers operate at the household level and are formally recognised by the state government, whereas the informal waste pickers and cart pushers (also operating at the household level) do not have formal recognition as they are not licenced to operate in the state. As in many other DCs, the informal waste pickers operate both at the household and urban scales, collecting recyclables from households, streets, and dumpsites under health-threatening conditions [98,99], which is one of the reasons why the state government regulates their activities. Although there have been deliberate efforts by the state government to permanently stop the operations of these informal operators [100,101], they continue to proliferate due to the critical service gaps in the collection of MSW, which remains a challenge in several DCs. It is therefore not surprising that after a 7-year ban, cart pushers have returned to Lagos streets, owing largely to the inefficiency of the licenced waste collectors [102]. The interactions between the LAWMA and its formal actors are characterised by challenges, particularly in areas of regulation and service consistency. This institutional complexity significantly impacts the effectiveness of the LRI and underscores the need for a multi-faceted approach to improving waste management and recycling outcomes in the city.

5.3. Data Collection and Analysis

A total of 137 questionnaires (Agege = 20; Eti-Osa = 107; Surulere = 10) were administered on different days at the three locations to both recyclers and households. The selection of survey participants employed the convenience sampling method. Convenience sampling is a non-probability technique where individuals are selected based on ease of access for the researcher, rather than being representative of the entire population [103]. At the outset of the field survey, contact details of active recyclers that could easily be surveyed were provided by the Central Head of Recycling of the LAWMA. Apart from households in the locations where the pilot scheme of the LRI was tested, other households surveyed were individuals that supply HSRs to some of the recyclers. As these individuals could also be easily reached, they were included in the sample. In Agege, 17 residents and 3 recyclers were interviewed. In Eti-Osa, 106 residents and 1 recycler were surveyed, while in Surulere, 7 residents and 3 recyclers were interviewed. After a review of the responses provided by the respondents, a total of 127 questionnaires were considered valid for the analysis, while 10 were deemed invalid. The basis for invalidation is that some questions considered critical to the analysis were unanswered. The questionnaire for the households fielded questions on respondents’ demographic data, awareness of the LRI and the Pakam App, types of HSR separated for collection, collection frequency, household rating of recycling, and what motivates them to recycle, among others.

5.3.1. Exploratory Data Analysis

Exploratory data analysis (EDA) was conducted on the data obtained from the recyclers to gain initial insights into the LRI’s operations. The analysis focused on key aspects of the RP, including the distribution of collectors across the three LGAs. EDA techniques were applied to examine waste bin distribution coverage, waste collection frequency, estimation of the RDR, and other relevant factors. Descriptive statistics and visual representations such as bar charts and pie charts were used to illustrate patterns in collector demographics, RRs, and operational efficiencies. This exploratory phase provided an overview of the recycling landscape in the studied LGAs, setting the foundation for more detailed statistical analyses and interpretation of the initiative’s effectiveness.

5.3.2. Machine Learning Modelling

Four ML models—logistic regression, random forest (RF), XGBoost, and CatBoost—were employed to build classification models for predicting household motivation for recycling based on the survey data [27,104]. LR is a linear model commonly used for binary classification tasks. It estimates the probability of an instance belonging to a particular class using the logistic function presented in Equation (2) [73,105]:

$$P(y=1\mid x)={\displaystyle \frac{1}{1+e^{-z}}}$$

(2)

where $z={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+\dots +{\beta }_n{x}_n$.

RF is an ensemble learning method that constructs multiple decision trees and combines their predictions. The final prediction is the average of individual tree predictions, which is represented by Equation (3) [71,106]:

$$f\left(x\right)={\displaystyle \frac{1}{n}} {\sum }_{i=1}^n{t}_i\left(x\right)$$

(3)

where $t_i\left(x\right)$ is the prediction of the $i-th$ tree.

XGBoost is a gradient boosting algorithm that builds decision trees sequentially, each tree correcting the errors of its predecessors. The objective function is a combination of a loss function and a regularisation term, as shown in Equation (4) [74]:

$$Obj \left(\theta \right)={\sum }_{\mathrm{i}=1}^{n}l\left({\mathrm{y}}_{\mathrm{i}},\widehat{y_i}\right)+{\sum }_{k=1}^K\Omega \left({\mathrm{f}}_k\right)$$

(4)

where $l$ is the loss function; ${\widehat{y}}_i$ is the predicted value; and $\Omega$ is the regularisation term.

CatBoost is another gradient boosting algorithm employed in this study to predict household motivation for recycling. It employs a novel, ordered boosting approach with the following loss function shown in Equation (5) [75]:

$$L={\sum }_{i=1}^{n}l\left(y_{i, }F\left(x_i\right)\right)+\Omega \left(\mathrm{F}\right)$$

(5)

where $l$ is the loss function; $F$ is the model; and $\Omega$ signifies the regularisation parameter.

Bayesian optimization (BO) was used to optimise the hyperparameters of these ML models. This approach employs probabilistic models to guide the exploration of the hyperparameter space, enabling efficient identification of the optimal configurations for improved model performance [107].

5.3.3. Evaluation Metrics

Given the unbalanced nature of the data employed in this study, the following evaluation metrics were used to assess the performance of the ML models: accuracy (Equation (6)), weighted recall (Equation (7)), weighted precision (Equation (8)), and weighted F1 score (Equation (9)). These metrics provided a more robust evaluation of model performance when dealing with imbalanced classes [73,108].

• Accuracy

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(6)

where TP is the true positives; TN is the true negatives; FP is the false positives; and FN is the false negatives.

2.

Weighted Recall

$$Weighted Recall=\sum w_i\times Recal{l}_i$$

(7)

where $w_i$ is the proportion of samples in class $i$ and $Recal{l}_i={\displaystyle \frac{T{P}_i}{TP{i}_i+F{N}_i}}$ for class $i$

3.

Weighted Precision

$$Weighted Precision=\sum w_i\times Precisio{n}_i$$

(8)

where $w_i$ is the proportion of samples in class $i$ and $Precisio{n}_i={\displaystyle \frac{T{P}_i}{TP{i}_i+F{P}_i}}$ for class $i$

4.

Weighted F1 Score

$$Weighted F1 Score=\sum w_i\times F{1}_i$$

(9)

where $w_i$ is the proportion of samples in class $i$ and $F{1}_i=2\ast {\displaystyle \frac{Precisio{n}_i\ast Recal{l}_i}{Precisio{n}_i+Recal{l}_i}}$ for class $i$.

These weighted metrics accounted for the class imbalance by incorporating the proportion of each class in the dataset. This ensures a more accurate and comprehensive evaluation of model performance across all classes.

5.3.4. Interpretability Using SHAP

To present the interpretability of the best-performing model and understand the factors influencing household recycling motivation in Lagos, SHAP values were utilised. SHAP is a game-theoretic approach that explains ML model outputs by calculating the contribution of each feature to a prediction. The SHAP value for a given feature $i$ is computed using Equation (10):

$${\varnothing }_j={\sum }_{S\subset N\backslash \left\{j\right\}}{\displaystyle \frac{k!(p-k-1)!}{n!}}[f(S \cup \{j\})-f(S)]$$

(10)

where n is the total number of features; $N\backslash \left\{j\right\}$ is a set of all possible feature combinations, excluding ${\varnothing }_j;$ S is a subset of $N\backslash \left\{j\right\};$ and $f(S \cup \left\{j\right\})$ is the prediction of the model, including features in $S$ and ${\varnothing }_j$ while $f$(S) is the prediction of the model with feature $S$.

5.3.5. Data Validation and Cross-Validation Approach

To ensure the reliability and validity of the model findings, a comprehensive data validation protocol was implemented. Data cleaning was performed to identify and handle missing values, inconsistent responses, and outliers. From the 137 questionnaires administered, 10 were invalidated due to these irregularities. Moreover, feature transformation was conducted to transform categorical variables into numeric representations suitable for ML algorithms while preserving their semantic meaning. The data validation process also included checks for multicollinearity among predictor variables to avoid redundancy in the feature set and ensure stable model performance

5.3.6. Addressing Sampling Biases

The convenience sampling method employed in this study may present inherent limitations regarding the representativeness of the sample. Several strategies were implemented to address potential sampling biases. First, k-fold cross-validation (k = 5) was employed during model training to ensure that performance metrics were not artifacts of a particular data split, thus providing a more robust assessment of model generalisability. Second, sample size sensitivity analysis was conducted by varying the training/test split ratios (30:70, 50:50, 70:30, and 90:10) to determine how model performance varied with different data partitions and to identify how it affected the stability of the model. Third, potential demographic biases were acknowledged, particularly the uneven distribution of respondents across the selected LGAs and socioeconomic factors. This geographical imbalance was considered during result interpretation, with findings explicitly framed within the context of these sampling limitations. These validation approaches collectively strengthen the reliability of our findings while acknowledging the constraints of the sampling methodology.

6. Conclusions and Recommendations

6.1. Conclusions

The findings from this research strongly support the hypothesis that the LRI has not achieved substantial landfill diversion rates due to operational challenges and insufficient household participation. In addressing the first research objective, the evaluation revealed a significantly underperforming RP with a low RDR of just 0.37%, falling dramatically short of the city’s ambitious 50% target. For the second objective, we successfully developed ML models to predict household recycling motivation, with the optimised CatBoost algorithm demonstrating the highest performance (accuracy and F1 score of 0.79). The third objective was fulfilled through SHAP analysis, which confirmed that operational factors, specifically Hct and Nob, are the most critical determinants of household recycling motivation, validating the operational challenges component of the hypothesis. By quantifying this performance gap and identifying key motivators for recycling participation through robust ML techniques, these findings address a critical gap in the literature on RIs in SSA where such evidence-based assessments are scarce. Although the analysis identified Hct and Nob as primary motivators of recycling specific to the LRI implementation, it should be noted that other researchers in SSA have identified additional factors such as household income, formal education, and age as general determinants of recycling behaviour. These results could serve as a benchmark for future investigations and policy formulations in other African cities grappling with similar WM challenges, potentially catalysing more effective and data-driven approaches to recycling and waste diversion across the region. Future research could expand data collection to cover a wider range of LGAs in Lagos, incorporating all waste diversion strategies, including reuse, composting, and waste-to-energy, to enable a more holistic assessment of the WDR in Lagos.

6.2. Recommendations

To improve the performance of the LRI beyond the current level, the LAWMA could make efforts to increase city-wide awareness of the programme through targeted public campaigns and engage frequently with the registered recyclers to understand and address their operational challenges. To further enhance the effectiveness of the initiative, policymakers could also implement financial incentives, such as subsidies or tax breaks for households and recyclers who participate actively in recycling. Also, investments in infrastructure, such as the provision of more recycling bins in high-density areas as well as the modernisation of waste collection fleets will improve the performance of the LRI. Additionally, the imposition of penalties for non-compliance with existing recycling regulations can boost participation rates, thereby ensuring the LRI meets its waste diversion targets. There is equally a need for the authority to focus on expanding the coverage of collectors across all the 20 LGAs to increase the collection rate of recyclables. These efforts would collectively contribute to enhancing the efficiency and effectiveness of the LRI, ultimately leading to improved WM practices across the Lagos metropolis. By addressing these recommendations and learning from the Lagos experience, cities across SSA and other developing regions could make substantial progress toward their waste diversion goals, thus contributing to improved environmental sustainability, public health, and economic opportunities for their residents. Moreover, the findings from this study could serve as a benchmark for future investigations and policy formulations in other African cities, potentially catalysing more effective and data-driven approaches to recycling and waste diversion across the region.