Enhancing Machine Learning Innovative Model for Waste Management: A Focus on Data Preprocessing and Labeling

Effective data collection is fundamental to building reliable machine learning models for waste classification. Previous studies have primarily utilized three types of data collection methods: primary, secondary, and tertiary sources. Primary data collection involves gathering images directly from real-world environments, often using cameras or sensors to capture waste in situ, which ensures relevance and context-specific accuracy [1, 23]. However, primary data collection can be resource-intensive and challenging to scale. Secondary data, on the other hand, is sourced from publicly available repositories like Kaggle or open-access image databases [24], offering a broader range of waste types and conditions. While secondary data is more accessible, it often varies in quality and consistency. Tertiary data collection, less common in waste classification studies, involves compiling insights from previous research or aggregating multiple datasets, providing additional context and enhancing model robustness by filling gaps in primary and secondary data [19, 25].

Preprocessing techniques are essential to prepare datasets for machine learning models and improve model performance by ensuring consistency and quality. Annotation and labeling, for instance, are widely recognized as critical steps in creating structured datasets; these techniques involve manually identifying and categorizing objects in images to enable accurate classification [26, 27]. Partitioning datasets into training, validation, and testing sets is another common practice, ensuring that models are assessed fairly and can generalize to unseen data. Normalization and standardization are also frequently applied to bring image data to a consistent scale, which helps models interpret pixel values uniformly across datasets. Data cleaning, including removing duplicates and handling missing values, is crucial for eliminating noise and ensuring dataset integrity [28]. Despite these advances, most studies employ these preprocessing techniques in isolation, which may limit the benefits that a comprehensive preprocessing pipeline could provide.

While traditional preprocessing techniques such as resizing, normalization, data augmentation, and annotation have been widely applied in waste classification [14-19], their implementation in dynamic real-world environments remains underexplored. Most studies rely on datasets captured under controlled lighting and simple backgrounds, limiting model robustness when deployed in practical settings characterized by variable illumination, cluttered scenes, and mixed waste piles  For instance, Zhang et al. [20] developed an MRS-YOLO model using a dataset collected under diverse lighting conditions (natural daylight, shadows, and artificial night lighting) and varied backgrounds (indoor, outdoor, cluttered, and simple), applying basic preprocessing steps such as resizing and normalization to improve generalization [7]. Similarly, utilized multi-source data (open datasets, social media, and environmental agencies) combined with targeted data augmentation, to handle rotations, lighting variations, and complex backgrounds in real-time waste detection. Further demonstrated that YOLOv8 models trained with augmentation for varying angles, positions, lighting, and resolution achieved higher robustness, yet still faced challenges with transparent and reflective materials in uncontrolled environments.

Despite these advances, a significant gap persists in the integrated application of multi-source data collection and comprehensive preprocessing pipelines tailored to highly variable waste scenarios. Existing works typically employ isolated preprocessing techniques or rely on single-source datasets, which fail to fully capture the ecological variability encountered in coastal or urban smart-city deployments .This study addresses these limitations by integrating primary (field-collected under uncontrolled conditions in Batam), secondary (public repositories), and tertiary (Google Images) data sources, followed by a rigorous preprocessing pipeline that includes unsupervised K-Means clustering for initial grouping, perceptual hashing for duplicate removal, label harmonization, normalization, and validated data augmentation. By annotating only high-quality subsets while preserving the full raw dataset for future semi-supervised learning, this approach not only enhances data diversity and quality but also provides a more practical and scalable foundation for real-world YOLO-based waste classification under dynamic lighting and mixed-background conditions.

Various machine learning models have been applied to waste classification, each with its own strengths and limitations. Traditional models like Random Forest and Support Vector Machine (SVM) are often used due to their interpretability and robustness with smaller datasets [9, 29]. However, these models may struggle with complex, high-dimensional image data, especially when waste types exhibit subtle visual differences. Deep learning models, particularly Convolutional Neural Networks (CNNs), have shown great promise in waste classification tasks because of their capacity to handle image data with complex patterns [23, 30]. Among CNN-based models, YOLO has gained popularity for real-time object detection and classification due to its speed and accuracy [31], making it well-suited for dynamic waste sorting applications. Nevertheless, deep learning models require extensive labeled data and computational resources, which can limit their feasibility in certain contexts.

Despite the significant progress in waste classification research, several gaps remain unaddressed. Many studies rely on limited data diversity, often using only primary or secondary sources, which can restrict the model’s ability to generalize across different waste types and conditions [1, 9, 19, 29]. Additionally, while preprocessing techniques like annotation and partitioning are applied, they are seldom integrated into a cohesive pipeline that addresses all aspects of data preparation. This fragmented approach can reduce model performance by leaving certain issues, such as class imbalance or data inconsistency, unresolved. Furthermore, while deep learning models like CNNs and YOLO have shown promising results, there is a lack of comprehensive studies comparing these models with traditional classifiers to assess their respective strengths in waste classification. This research aims to address these gaps by developing a multi-source data collection strategy, implementing an integrated preprocessing pipeline, and evaluating the effectiveness of YOLOv8 against traditional models, providing insights into the optimal approach for automated waste classification.

4.1 Model selection

In this study, a range of machine learning models was selected to evaluate the impact of the data collection and preprocessing techniques on waste classification accuracy. The primary model employed was YOLOv8, a cutting-edge object detection model designed for real-time applications that require both high speed and accuracy. YOLOv8, the latest iteration in the YOLO series [32-34], is particularly well-suited for tasks involving the detection and classification of multiple objects within an image, such as waste sorting. Its architecture allows it to quickly process images, detect objects within bounding boxes, and classify them based on the learned features, making it ideal for scenarios like waste management where rapid and accurate classification is critical. YOLOv8’s ability to handle complex object detection with high efficiency and accuracy positions it as the core model for this study.

To provide a comprehensive analysis and benchmark YOLOv8’s performance, several traditional machine learning models were also included in the study. These models, which served as a point of comparison, included Random Forest, SVM, Multi-layer Perceptron (MLP), Logistic Regression, and Naive Bayes. Each of these models offers unique advantages for classification tasks. Random Forest is an ensemble learning technique known for its robustness and its ability to manage diverse datasets by combining multiple decision trees [35]. SVM is effective in high-dimensional spaces and can be particularly useful for tasks where the data are not linearly separable [36]. MLP, a type of neural network, can capture complex patterns in data but typically requires substantial computational resources and a large dataset to perform well [37]. Logistic Regression is a straightforward yet effective baseline classifier for binary and multi-class tasks [38], while Naive Bayes is a probabilistic model that assumes feature independence, offering simplicity and interpretability [39].

The selection of YOLOv8 as the primary model and the inclusion of traditional models for benchmarking were strategic. YOLOv8 was chosen due to its advanced capabilities in handling object detection tasks with high speed and accuracy [34], which are essential for real-time waste classification systems. However, to assess the model’s true effectiveness, it was necessary to compare its performance against traditional machine learning models. This comparison allows for a deeper understanding of whether the added complexity and resource demands of a model like YOLOv8 are justified or if simpler models like Random Forest or Logistic Regression can achieve comparable results with fewer computational requirements. By evaluating these models side-by-side, this study provides insights into the relative strengths of each approach, helping to determine the optimal model choice for scalable waste management applications.

4.2 Training process

YOLOv8, the primary model in this study, was trained through a structured process designed to maximize its effectiveness in object detection and classification. Training began with model initialization, where parameters such as learning rate, batch size, and model depth were configured to suit waste classification. These parameters were initially set based on standard recommendations and fine-tuned throughout the training to adapt to the dataset’s specific characteristics. The model utilized the Adam optimizer, an efficient algorithm for deep learning tasks [40], to minimize the loss function by continuously adjusting weights based on the difference between predicted and actual labels, thereby improving detection and classification accuracy over iterations.

In addition to YOLOv8, traditional models such as Random Forest, SVM, MLP, Logistic Regression, and Naive Bayes were trained as benchmarks. Each model’s training process was adapted to its specific algorithm. For example, Random Forest involved constructing multiple decision trees and aggregating their predictions, while SVM focused on finding optimal hyperplanes to separate waste categories in a high-dimensional feature space. MLP employed backpropagation to adjust weights across layers, capturing complex data patterns, whereas Logistic Regression and Naive Bayes served as baseline models requiring fewer computational resources. To objectively evaluate and compare these models, various performance metrics were applied, including accuracy, precision, recall, F1-score, and mean Average Precision (mAP) for YOLOv8. These metrics collectively provided a comprehensive view of each model’s performance, with mAP specifically reflecting YOLOv8’s object detection accuracy. This comparison offered insights into the relative strengths and limitations of traditional models versus a modern object detection model like YOLOv8.

6.1 Key findings

Beyond the raw performance numbers, the relationship between metrics provides deeper insight. YOLOv8 achieved a strong mAP@0.5 of 0.8000 while maintaining a balanced F1-score of 0.7557, indicating robust precision-recall trade-off. However, the gap between precision (0.7304) and recall (0.7857) suggests the model is slightly more conservative in predictions, which is advantageous for waste management (minimizing false positives in sorting lines) but still leaves room for recall improvement on minority classes.

The comparative analysis underscores a clear advantage of deep learning architectures over traditional methods when dealing with complex, real-world image data. Traditional models not only achieve lower overall metrics but also show pronounced degradation under challenging conditions such as occlusion and high feature density—conditions that are prevalent in practical waste management deployments.

6.2 Broader implications

The findings of this study have direct and significant implications for real-world waste management systems, particularly in automated sorting facilities and smart-city initiatives. While the statistical performance of YOLOv8 (mAP@0.5 = 0.8000, F1-score = 0.7557) is promising, its practical value becomes clearer when translated into operational outcomes. For example, false classifications between recyclable and non-recyclable items-such as mistaking transparent plastic for glass or plastic wrappers for metal foil-can lead to cross-contamination in recycling streams [41, 42]. In a typical municipal sorting facility processing 500 tons of waste per day, even a 5-10% increase in contamination (a realistic scenario based on the observed misclassification rates in our test set) could raise downstream processing costs by 15-25% and reduce the recovery rate of valuable materials by up to 12%, according to industry benchmarks [43-45].

Furthermore, the model’s relatively lower performance on rare categories (e.g., glass, cardboard, tissue, and food plastic wrappers, which comprise < 8% of the annotated dataset) highlights a critical deployment challenge. In actual waste management facilities, these minority classes often represent high-value recyclables or hazardous materials. Missing them would result in lost revenue from recycling and potential environmental harm. Deploying the current model “as-is” would therefore require targeted improvements, such as class-weighted loss functions, additional primary/IoT data collection for rare categories, or hybrid sensor fusion (visual + near-infrared) to achieve the >95% accuracy threshold typically demanded by industrial-scale automated sorting lines.

These practical considerations demonstrate that the multi-source dataset and comprehensive preprocessing pipeline developed in this study not only improve statistical metrics but also provide a more robust foundation for real-world applications, bridging the gap between laboratory performance and operational efficiency in dynamic urban waste management environments.

Additionally, a well-structured preprocessing pipeline is vital for maintaining data quality, which directly impacts the efficacy of waste classification models. Preprocessing steps such as annotation, partitioning, cleaning, normalization, and augmentation ensure that the dataset remains consistent, comprehensive, and reflective of real-world scenarios, as highlighted by studies in image classification that advocate for robust preprocessing to enhance model performance [46, 47]. For example, accurate annotation and balanced data partitioning enable models to distinguish subtle differences between waste categories, reducing misclassification rates that could otherwise disrupt automated sorting processes. Normalization and augmentation further improve adaptability by accounting for variations in lighting and orientation, common in waste sorting facilities. This systematic preprocessing approach helps smart cities achieve higher model performance, contributing to more reliable and efficient automated sorting systems. In a broader urban sustainability context, such intelligent waste management solutions minimize reliance on manual sorting, lower operational costs, and enhance recycling efficiency, supporting environmental goals and aligning with the data-driven vision of smart cities for sustainable urban ecosystems.

In practical smart-city applications, these results imply that YOLOv8 can be integrated into real-time waste sorting robots or conveyor-belt systems with high reliability. However, the identified confusion between transparent materials underscores the importance of combining visual detection with complementary sensors (e.g., near-infrared spectroscopy) in future hybrid systems to achieve >95% accuracy required for industrial-scale deployment.

6.3 Relation to existing literature

The findings of this study build upon and extend the existing body of research on waste classification using machine learning, demonstrating both advancements and alignment with previous studies. Previous research has consistently highlighted the significance of data quality and diversity in enhancing model accuracy for waste classification tasks. For instance, studies by Pitakaso et al. [1] and Ren et al. [6] emphasized the benefits of utilizing primary and secondary datasets to increase the range of waste types available for model training. This study aligns with those findings by adopting a multi-source data collection strategy that includes primary, secondary, and tertiary data, thereby enhancing the dataset’s diversity and enabling the model to generalize more effectively across different waste types and conditions. The success of YOLOv8 in this study reinforces the importance of comprehensive datasets, validating earlier conclusions about the role of diverse data in improving model performance.

In addition to data diversity, this study’s use of a robust preprocessing pipeline resonates with established practices in the field. Previous studies, such as those by Simoni et al. [48] and Kunft et al. [49], have shown that proper preprocessing steps-including annotation, normalization, and data cleaning-are essential to reduce noise and increase the accuracy of classification models. This study corroborates those findings by demonstrating how a structured and integrated preprocessing pipeline contributes to YOLOv8’s strong performance, particularly in distinguishing between visually similar waste types. Unlike prior studies that often applied preprocessing techniques in isolation, this research adopted a comprehensive pipeline, systematically combining multiple techniques. This approach represents a significant advancement in preprocessing practices, showing that an integrated pipeline can enhance model robustness and adaptability in complex classification tasks like waste sorting.

The choice of YOLOv8 as the primary model also marks an advancement over traditional models frequently used in earlier studies. Research by Li et al. [50] and Shad et al. [51] predominantly employed conventional models such as Random Forest and SVM, which, although effective to a degree, have limitations in handling high-dimensional image data. These models generally struggle with the complexity and variability inherent in waste classification. By contrast, YOLOv8, with its advanced object detection capabilities, achieved significantly higher accuracy, precision, and recall in this study. This improvement highlights the potential of deep learning models, particularly YOLO-based architectures, for real-time waste classification, an essential requirement for dynamic urban waste management systems. The results suggest that modern object detection models like YOLOv8 are better suited for complex image-based classification tasks, marking a shift in best practices for automated waste management research.

This study contributes to the growing field of smart waste management by aligning with established research on data diversity and preprocessing while also demonstrating significant advancements in methodology and model choice. The findings underscore that combining a diverse data collection strategy with an integrated preprocessing pipeline and state-of-the-art deep learning models can result in substantial improvements in classification accuracy. These insights provide a valuable foundation for future research to refine and expand upon these methodologies, potentially setting new standards for automated waste sorting solutions in smart cities.

This study extends the literature by moving beyond isolated model-centric improvements to a holistic data-centric approach. While previous works have explored individual preprocessing steps or single-source datasets [14-19], few have systematically integrated primary, secondary, and tertiary sources with a full preprocessing pipeline explicitly designed for dynamic environments. The combination of multi-source data diversity and rigorous preprocessing (including quality filtering and validated augmentation) enables superior handling of real-world challenges such as variable lighting and cluttered backgrounds-challenges that remain prominent limitations in many YOLO-based studies [35]. Thus, this research contributes a novel, reproducible framework that bridges the gap between controlled laboratory datasets and practical smart-city waste management applications.

6.4 Limitations

While this study highlights the benefits of an integrated data collection and preprocessing pipeline for waste classification, several limitations should be addressed. One major limitation is the reliance on secondary data sources, such as Kaggle, for a significant portion of the dataset. Although secondary data sources provide a large volume of images, they often lack context-specific details, potentially limiting the model’s applicability to real-world scenarios where waste types and conditions may vary [18]. Additionally, secondary datasets frequently exhibit inconsistencies in labeling standards and image quality, which required extensive data cleaning to ensure compatibility across sources [52]. To overcome this limitation, future studies could integrate more primary data collected from local environments to improve the dataset's relevance and accuracy for specific waste management applications [19, 53].

Manual annotation posed another challenge in this study due to its time-consuming and labor-intensive nature. The manual labeling process required detailed identification of each waste type within images to maintain accuracy. However, this process is prone to human error, especially when distinguishing between visually similar waste items like plastic wrappers and metal foils, which may lead to mislabeling and ultimately impact model training and performance [54]. The use of AI-assisted annotation tools or semi-supervised learning techniques could alleviate some of these challenges, providing a more efficient and consistent method for annotating large datasets [55].

The annotated dataset suffers from severe class imbalance, with plastic_bottle accounting for 31.7% of training samples while five minority classes each represent less than 8%. This imbalance likely drives the overall accuracy of 78% and mAP of 80%, as evidenced by the per-class metrics, where recall on metal, cardboard, and tissue is substantially lower than on dominant classes. Although stratified splitting and data augmentation were applied, the impact remains significant and limits generalizability. Future work will address this through class-weighted loss, oversampling of minority classes, or semi-supervised learning on the remaining 8,953 raw images.

These limitations suggest pathways for further improvement, emphasizing the need for targeted primary data collection, automated annotation tools, and strategies to address dataset balance. By implementing these improvements, future studies could enhance the robustness, efficiency, and generalization capabilities of waste classification models, supporting more effective applications in automated waste management [56].

Sample Size Limitation (New): A critical limitation of this study is the disconnect between raw data collection and annotated data availability. While we collected over 10,000 images, only 1,093 (10.9%) were fully annotated with bounding boxes due to the labor-intensive nature of manual annotation. This small annotated sample size, particularly for minority classes like glass (n = 54) and cardboard (n = 52), limits the statistical power and generalizability of our findings. Performance metrics for categories with <100 training samples should be interpreted with caution, as they are more susceptible to sampling variability [57]. Future work should prioritize: (a) scaling up annotation efforts through crowdsourcing or AI-assisted tools, (b) employing semi-supervised learning to leverage the unannotated 8,953 images, or (c) using active learning to identify the most informative samples for annotation.

Furthermore, while preprocessing and multi-source data improved generalization, residual errors on transparent and reflective materials indicate that current normalization and augmentation strategies are insufficient for handling optical properties unique to certain waste types. Future optimization could explore attention mechanisms focused on texture and material properties or domain-specific data augmentation (e.g., synthetic rendering of translucent objects).

A critical limitation of this study is the relatively small size of the annotated dataset (1,093 images) compared to the raw collection of 10,046 images. Although this was unavoidable due to the high cost of high-quality bounding-box annotation, it limits statistical power, especially for minority classes. Future work will address this through semi-supervised learning on the full raw dataset.

Data collected via IoT-like field devices introduced additional variability (lighting changes, uneven resolution, and positional differences) that increased the difficulty of model training. Although these challenges were mitigated through targeted preprocessing and augmentation, they still contribute to the relatively modest performance on certain minority classes and highlight the need for future sensor calibration or multi-modal fusion in operational IoT systems.

Correction to Earlier Reporting: We acknowledge that the initial manuscript did not clearly distinguish between raw dataset size (10,046) and annotated dataset size (1,093). This has been corrected in the revised manuscript, with Table 1 now explicitly labeled as "Raw Dataset" and Table 2 as "Annotated Dataset." All reported model performances are based solely on the annotated dataset.