Smart waste management and classification system using advanced IoT and AI technologies

Technology background

Related work

In an age characterized by the rapid modernization of all societal sectors, technology has become an integral aspect of daily life for individuals. The profound alterations in human lifestyles, propelled by technological advancements, are observable as modern technology permeates every facet of human existence. Cities endowed with cutting-edge facilities and technology are preferred, leading to a continual surge in urban population, accompanied by both advantages and disadvantages. The increasing congestion in growing urban areas brings about worries concerning health, safety, and environmental issues, which encompass aspects such as medical services, security, privacy, and transportation (Javed et al., 2022).

Solid waste management emerges as a notable challenge in contemporary urban areas, involving intricate processes encompassing to collection and transferring at particular dump area, classification, and recycling the segregated solid waste. The heightened utilization of natural resources has resulted in alarming levels of waste, posing threats to human health and the environment. Despite the awareness among the educated population in urban areas regarding the environmental impact of waste, a significant portion disposes of waste without proper sorting and recycling.

Recycling assumes a central role as evidenced by studies indicating that nearly 0.75% of solid domestic waste can be recycled, presenting economic benefits and environmental sustainability. Countries worldwide, including America, China, Japan, Canada, Korea, Russia, Italy, Malaysia, Saudi Arabia, and Qatar, actively engage in developing smart waste management systems (Fidje, Haddara & Langseth, 2023). Previous endeavors involve techniques implemented in cities like St. Petersburg, Russia, utilizing technologies such as wireless sensor networks (WSNs), radio frequency identification (RFID), sensors, and actuators (Ogarkov, 2019; Vishnu et al., 2021).

Adedeji & Wang (2019) introduced a novel waste management system tailored for smart cities, harnessing the capabilities of the IoT to overcome limitations inherent in conventional waste management systems. The proposed framework consists of two distinct types of end sensor nodes: public bin level monitoring units (PBLMUs) designed for monitoring public bins, and home bin level monitoring units (HBLMUs) intended for residential areas. Subsequently, they process this information and transmit it to a centralized monitoring station for storage and analysis. The experiments were conducted to assess the power consumption of a PBLMU and estimate its life expectancy under hypothetical conditions, providing valuable insights into the system’s practicality and efficiency. The data extracted from the IoT system, including trash bin details such as unfilled level and geolocation coordinates, can be employed to create a geographic information system (GIS). Additionally, ML algorithms can be applied to determine optimal routes for waste collection trucks as well as segregation of the waste.

In the domain of waste classification, various image classification techniques, including support vector machine (SVM), CNN, and ResNet-50, have been employed (Lilhore et al., 2024). The rise of deep learning models, particularly CNN architectures like VGG16 (Ahmed et al., 2023) and VGG19 (Vaidya & Paunwala, 2019; Bansal et al., 2023), has demonstrated increased effectiveness for object detection and classification compared to traditional methods. The incorporation of IoT technologies into waste management systems has given rise to the development of intelligent and automated waste management solutions in smart cities, thereby enhancing efficiency and sustainability.

Leading to IoT and emerging technologies, Shahab & Anjum (2022) worked on sustainable waste management and proposed a novel AI and IoT based multi-path (MP) convolutional neural network model. The proposed study was particularly for delhi region in India. They defined the category into two major classes including waste and non-waste. Further waste category is dumped at specific region and implement waste classification. According to the authors, the proposed MP-CNN model achieved 98% accuracy in waste detection while segregating process, however it was tested in short volume data with just 600 images. The same mechanism was proposed by Krishna & Sharma (2023) with improved datasets containing 1,150 images. They used multiple classifiers including VGG-16, ResNet-16 etc., but their performance was not up to the mark and achieved 93% accuracy. To improve the waste collection process, Sheng et al. (2020) implemented LoRa protocols for better communication while collecting waste from smart dustbins. According to authors the communication channels should be stronger to provide the correct information and alerts when a bin is filled and supposed to be dumped for further segregation. Indeed the proposed study achieved remarkable outcome and improved the waste collected process using advanced IoT technologies, however, waste segregation was not addressed in their proposed study. Below is a comparative analysis of existing state of the art proposed studies presented in Table 1.

Though several techniques have been proposed over sustainable waste collection and segregation process as discussed above, however, it has been observed that the implementation ibn real environment is still a challenge being face worldwide. To cope this challenge, this study proposed a novel scheme that deal both waste collection mechanism as well as segregation discussed in following section.

iSSWMs elements

Smart bins

The smart dustbin, equipped with advanced sensors such as weight and volume sensors, revolutionizes waste collection by efficiently monitoring and managing waste levels. In this study, we have connected four bins including glass bin, plastic bin, metal bin, and trash bin (used for waste other than selected waste types). The controller receives the sensory information continuously from all the connected bins, and forward to cloud server along with bin location information for further processing and storage. Figure 1 presents the working principle of smart bin as follows:

iSSWMs workflow

The proposed iSSWMs framework works in two phases. First phase is regarding the waste collection and data analytics. Further phase two deals with waste segregation using deep learning. Figure 2 shows the block diagram of proposed iSSWMs architecture. The detail workflow of iSSWMs architecture is as follows:

Phase 1: Intelligent bin monitoring and cloud-based analytics

In the first phase of the smart waste management system, the integration of cutting-edge technologies is strategically employed to optimize efficiency and promote sustainability. The process begins with the deployment of four intelligent bins, each designated for specific waste types: glass, plastic, metal, and organic waste. These smart bins are not merely receptacles; they are equipped with advanced sensors that continuously monitor critical parameters such as bin size, waste quantity, and precise bin location. The intelligence of these bins comes to the forefront when one of them reaches its capacity. At this point, the bin autonomously transmits a comprehensive set of data to the cloud. This data transmission includes real-time information about the bin’s size, the quantity of waste it contains, and its geographic location. The cloud-based database acts as the nerve center, swiftly processing and analyzing the incoming data. Subsequently, the relevant information is relayed to a dedicated mobile application, providing instant accessibility to all concerned authorities involved in waste management. Upon reception of the notification through the mobile application, designated personnel are promptly alerted to the specific location where waste needs to be collected. The collected waste is then systematically transported to a centralized dump area for further processing. This phase exemplifies the seamless coordination between intelligent bins, cloud-based analytics, and mobile applications, fostering a responsive and efficient waste collection process. A detailed procedure of waste collection and segregation is presented in Algorithm 1.

Phase 2: Waste segregation (CNN)

The second phase of the smart waste management system delves into the intricacies of waste segregation, aiming to streamline the process through advanced automation. The waste collected in the central dump area is now subjected to a systematic segregation procedure. The waste items are conveyed on a sophisticated belt system, enhancing the efficiency of the overall process. As each item moves along the conveyor belt, a fixed camera captures detailed images of the waste. The integration of deep learning techniques, particularly CNN models, is instrumental in image recognition. The CNN models analyze the captured images, swiftly identifying the type of waste based on predetermined categories, such as glass, plastic, metal, or organic waste. The analysis information is forwarded to microcontroller which is controlling the attached three actuator pump placed on conveyor belt. Once the waste item start moving on conveyor belt, the relevant actuator pump push the waste item to perspective bin through connected slide. In this scenario, if the waste item is not recognized appropriately, the waste item is dropped in trash bin which is basically an irrelevant waste from the defined categories. This systematic and smart process segregate the waste items accurately with less power consumption and almost zero manpower. This automation significantly contributes to the promotion of recycling and environmentally responsible waste management practices. Figure 2 shows the block diagram of proposed iSSWMs.

For this study, the VGG-19 model was chosen due to its well-established performance in image classification tasks. VGG-19, originally trained on the large-scale ImageNet dataset, consists of deep convolutional layers capable of extracting hierarchical features from input images. Instead of training a deep learning model from scratch, which would require extensive computational resources and a larger dataset, transfer learning was employed by utilizing the pre-trained weights of VGG-19 and adapting them to the waste classification task.

The integration process of the proposed iSSWMs model ensures seamless connectivity between smart bins, cloud-based analytics, and automated waste segregation, making it highly scalable across diverse environments. The system’s modular design allows for easy deployment in urban, industrial, and commercial settings by adapting the number of smart bins and sensor configurations based on waste generation levels. The ThingSpeak platform enables real-time data processing and decision-making, facilitating effortless integration with existing waste management infrastructures. The cloud-based storage (Firebase) ensures scalability by efficiently handling data from multiple locations and providing remote access to stakeholders via mobile applications. Furthermore, the use of deep learning-based waste recognition, specifically VGG-19, enhances adaptability in different waste categorization scenarios, ensuring robustness even with varying waste compositions. The automated segregation process, driven by microcontroller-controlled actuator pumps, minimizes manual intervention and can be scaled by adding more conveyor lanes or actuator units for high-throughput environments. These design choices collectively support the system’s integration potential and scalability, ensuring efficient waste management across diverse operational landscapes.

Datasets description

This research employs the TrashNet dataset, accessible on GitHub, comprising around 3.5 GB of manually gathered data. The dataset encompasses six classifications including metal, paper, glass, cardboard, and trash, encompassing 99% of recyclable materials. At present, it encompasses 2,527 waste images, each with dimensions of 512 × 384 pixels, adjustable via resizing. These images were captured with items positioned on a white poster board under room illumination or natural light. Figure 7 showcases representative samples from each category within the dataset.

Data preprocessing: To enhance the quality of data input and improve model performance, we applied the following preprocessing steps:

• Resizing: All images were resized to 224 × 224 pixels to standardize input dimensions for VGG-19, ResNet, and MobileNet V2.

• Normalization: Pixel values were scaled to the range [0,1] to ensure numerical stability and accelerate convergence during training.

• Grayscale to RGB conversion: If an image was in grayscale, it was converted to a three-channel RGB format to match the expected input format of deep learning models.

• Noise reduction: We applied Gaussian smoothing to reduce unwanted noise without affecting important waste features.

Data augmentation techniques: To address dataset limitations and improve model generalization to real-world conditions, we implemented a data augmentation pipeline that artificially expands the dataset by introducing variations commonly encountered in waste collection environments:

• Random rotation (±30°): Simulates different orientations of waste objects.

• Brightness and contrast adjustments (±20%): Mimics varying lighting conditions.

• Horizontal and vertical flipping: Increases diversity by mirroring waste items.

• Occlusion simulation (random erasing): Introduces random occlusions to make the model robust against partial visibility.

• Background randomization (CutMix and MixUp): Combines different waste items to simulate complex real-world waste piles.

Handling class imbalance: TrashNet, like most real-world waste datasets, exhibits class imbalance, where certain waste categories (e.g., plastic and paper) have significantly more samples than others (e.g., metal and glass). To mitigate this, we adopted the following strategies:

• Weighted loss function: We used a class-weighted cross-entropy loss, where underrepresented classes were assigned higher weights to ensure balanced learning.

• Oversampling & synthetic data generation: We applied Synthetic Minority Over-sampling Technique (SMOTE) to generate synthetic samples for underrepresented waste categories.

• Batch balancing: During training, we implemented a stratified sampling approach to ensure each batch contained a proportionate mix of waste types, preventing the model from being biased toward majority classes.

For this research, a portion of the dataset focusing on plastic, metal, glass, and trash was employed. This subset was divided into two parts: one for training, and the other for testing, with proportions of 85% and 15%, respectively. However, to assess the model in real-time, the entire dataset was used for training, and testing was performed on images obtained in real-time. Table 3 presents the distribution of samples across categories in each subset, revealing variations in sample quantities per category.

Performance evaluation metrics

Results

This section presents the evaluated results of our proposed model, and comparison with existing state-of-the-art studies. We observe that our proposed iSSWMs reveals significant advancements for solid waste classification. Existing studies have employed architecture like ResNet, SVM, VGG-16, ResNet, AlexNet, DenseNet, MobileNet, and RCNN, focusing on different types of solid waste such as glass, metal, paper, plastic, and recyclable waste. While these methods achieved respectable accuracies ranging from 84% to 95%, the iSSWMs, utilizing VGG-19, ResNet, and MobileNet V2 architectures, demonstrated remarkable performance improvements across multiple waste types, including glass, plastic, metal, and general trash. Specifically, the iSSWMs achieved accuracy of 99.7%, 97.2%, and 98.18% for VGG-19, ResNet, and MobileNet V2, respectively, showcasing their superior effectiveness in solid waste classification tasks compared to existing methodologies as shown in Table 4.

Further we evaluated average precision against each waste item, and observed notable differences emerge. In previous studies (Adedeji & Wang, 2019; Ahmed et al., 2023; Vaidya & Paunwala, 2019), the AP generally increased with the number of epochs, reaching maximum values ranging from 79% to 92%. However, the proposed iSSWMs model consistently outperformed these existing techniques across all epochs and waste items (glass, plastic, and metal). As the number of epochs increased, the AP of the iSSWMs model exhibited substantial improvements, surpassing 99% with just 80 epochs and reaching an exceptional accuracy of 99.7% with 100 epochs as shown in Figs. 8–10, respectively.

Similarly, we evaluated recall measures the ability of a model to correctly identify all relevant instances (true positives) out of all actual positives. We also evaluated F1-score which is the harmonic mean of precision and recall. F1-score is useful when you need to strike a balance between precision and recall as shown in Table 5. It is especially valuable in situations where both false positives and false negatives are important, and the class distribution is imbalanced. Recall and F1-score are calculated as in Eqs. 5 and 6 respectively:

(5) $$Recall={\displaystyle \frac{Truepostive\left(TP\right)}{Truepostive\left(TP\right)+FalseNegtives\left(FN\right)}}$$

(6) $$F1=2X{\displaystyle \frac{precisonXrecall}{precison+recall}}$$

Further, we evaluate the accuracy and loss rate of our proposed iSSWMs model while training and testing. Figures 11, and 12 show the fitting process of iSSWMs model, displaying the accuracy and loss curves. Both graphs indicate the absence of overfitting, as the test curves consistently show an uptrend in accuracy or a downtrend in loss, mirroring the behavior of the training curves. This indicates the model’s effective learning in accurately classifying the data, extracting distinctive features across various stages, and utilizing them for class identification during both training and testing phases. Furthermore, it presents robustness against overfitting to the predominant class (middle), notwithstanding the intrinsic class imbalance within the dataset.