LPWAN Smart Waste Bin with On-Device AI Trained on Synthetic Data

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 1 LPWAN Smart Waste Bin with On-Device AI Trained on Synthetic Data Pavel Tr´ avn´ ıˇ cek Member, IEEE, V´ aclav Neˇ zerka Abstract—The construction industry is a major contributor to global solid waste, yet circular economy initiatives are often impeded by inefficient logistics and improper sorting. Existing monitoring solutions, typically reliant on single-point distance sensors, lack the granularity to identify waste composition, which is essential for effective valorization. This work proposes an energy-efficient, image-based smart bin system enabled by Low Power Wide Area Networks (LPWAN) that utilizes ondevice Artificial Intelligence (AI) to simultaneously estimate fill levels and classify waste materials. To address the scarcity of labeled field data, a synthetic data generation strategy using generative AI was employed to create photorealistic training datasets. A lightweight MobileNetV2 model was optimized via quantization and deployed on an ESP32 microcontroller. The system architecture prioritizes energy conservation by performing inference at the edge and transmitting only compact results, reserving full image transmission for a closed-loop active learning pipeline. Energy profiling demonstrated that on-device inference drastically reduces active radio time compared to raw image streaming, significantly extending battery life. The work validates the feasibility of Edge AI for scalable construction and demolition waste monitoring and highlights the potential of synthetic data to overcome data scarcity bottlenecks. Index Terms—IoT, skip bins, machine learning, synthetic images, edge computing THE construction industry plays a pivotal socio-economic role globally, generating approximately 25% of the global GDP and employing 7% of the population [1]. Within the European Union (EU) alone, the sector employed 18 million people in 2020 [2]. However, this economic contribution comes at a significant environmental cost. The sector is a massive consumer of raw materials and a prolific producer of waste. Globally, the construction industry is estimated to consume over 30–40% of all extracted natural resources [3], [4], generate approximately 25–40% of total solid waste [5], and emit up to 25% of anthropogenic CO2[6]. In 2020, the production of construction and demolition waste (CDW) in the EU was estimated at 747.3 million tons, equating to approximately 1685 kg per capita. To pursue sustainable development, managing waste prudently and cost-effectively while adopting circular economy principles is imperative [7], [8]. Although the EU maintains a high CDW recovery rate of approximately 90%, the majority is downcycled rather than recycled for high-value applications. Rapidly developing nations face even greater challenges; for instance, China produces approximately 2 billion tons of CDW annually, surpassing the combined output of all EU member P. Tr´ avn´ ıˇ cek is with the Faculty of Civil Engineering, Czech Technical University in Prague, Th´ akurova 7, 166 29 Praha 6, Czech Republic (e-mail: [email protected]). V. Neˇ zerka is with the Faculty of Civil Engineering, Czech Technical University in Prague, Th´ akurova 7, 166 29 Praha 6, Czech Republic (e-mail: vaclav[email protected]). states [9]. The primary obstacle to the valorization of crushed CDW in high-value applications, such as concrete manufacturing, is improper sorting [10]. Studies suggest that automated classification of CDW materials holds the greatest potential for promoting recycling and reuse, significantly reducing sorting costs [11], [12]. Despite the urgency, the tracking of waste skip-bins in CDW logistics has not become standard practice. Where tracking solutions exist, they rarely monitor the bin’s contents or fill level with sufficient granularity. Traditional monitoring implementations typically rely on single-point distance sensors, such as ultrasonic, LiDAR, or Time-of-Flight (ToF) modules. For example, Patel et al. [13] proposed a solar-powered smart bin utilizing ultrasonic sensors to measure trash levels and trigger servo motors for lid operation. Similarly, Lakhouit [14] reviewed IoT-enabled smart bins that use sensors to detect fill levels and odors, alerting collection services only when necessary to optimize routing. While these solutions represent a significant advancement over static collection schedules, they suffer from a limited field of view and, crucially, an inability to identify the specific waste material type, which is essential for CDW valorization. The integration of Artificial Intelligence (AI) and the Internet of Things (IoT) offers a transformative pathway to address these limitations. Atofarati et al. [15] highlight that AI and IoT can enhance operational efficiency and minimize environmental harm through predictive waste volume sensing and automated sorting. However, the architecture of such systems is critical. In many existing implementations, raw data is transmitted to centralized cloud servers for processing. Firouzi et al. [16] argue that relying solely on cloud computing for AI-driven IoT generates significant challenges regarding latency, bandwidth, and energy efficiency. Transmitting raw images from waste bins to a central server is impractical in low-power IoT environments due to bandwidth constraints and the high energy cost of continuous transmission. As fleet sizes scale, this centralized model results in unsustainable data costs and storage overhead. To resolve these bottlenecks, there is a paradigm shift toward “Edge AI,” where intelligence is distributed to the device level. Khadam et al. [17] identified that while cloud deployment is common, edge deployment is increasingly dominant in applications requiring real-time pattern recognition and reduced latency. By shifting analysis to the edge, devices can transmit only compact, actionable inference results rather than raw data streams. This approach aligns with the need for energy-efficient M2M (Machine-to-Machine) routing protocols that optimize network resources by selecting the most efficient nodes for data handling [18]. IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 2 Edge Node (ESP32 + NB-IoT) Image Capture & Sensing Edge Inference (INT8 TFLite) TCP Transmission (NB-IoT) Power management (Deep Sleep) Confidence Thresholding Model Deployment Offline Training & Optimization Base Image Acquisition Synthetic Data Gen. (GenAI) Data Augmentation & Training Quantization Input: Real-world feedback Real-world Data Feedback Cloud Infrastructure Stream Processing (Node-RED) Time-series Storage (InfluxDB) Fleet Dashboard (Web) Device Management Alerting system Telemetry + Image (TCP) Fig. 1. System architecture overview illustrating the closed-loop methodology: (left) offline synthetic training pipeline, (center) on-device edge inference and power management, and (right) cloud-based telemetry and active learning feedback loop. This work proposes an image-based approach to simultaneously determine fill level and material type, utilizing ondevice inference, building upon our initial system architecture presented in [19]. Optimization at the source is critical for the scalability of IoT deployments in battery-operated ecosystems. While on-device inference for a 320×240px image requires approximately 0.3 seconds, transmitting the same raw data over Low Power Wide Area Networks (LPWAN) can take upwards of 30 seconds. This reduction in active radio time allows the device to return to a low-power sleep state significantly faster, extending operational battery life. A significant hurdle in developing such AI models for CDW is the scarcity of labeled field data. To overcome this, we employed a synthetic data generation strategy, an approach successfully utilized in data-critical domains such as medicine [20]. Base images of empty skip bins, captured under varied lighting conditions, were generatively augmented to simulate diverse material types and fill states. This synthetic dataset serves as the foundation for the initial “Stage 0” training presented in this work. It must be noted that due to the limited availability of real-world datasets, the current model represents an early developmental phase requiring further refinement. Consequently, the primary contribution of this study is the comprehensive, open-source hardware and software framework designed to bootstrap this data collection and inference process. This paper details the design, synthetic training, and pilot deployment of the LPWAN smart bin system, laying the groundwork for a scalable solution that will mature through continuous training on future field-collected data. I. METHODOLOGY To enable scalable waste monitoring in resource-constrained environments, we propose a methodology centered on datacentric AI and edge computing. The system architecture, illustrated in Fig. 1, integrates synthetic data generation with an optimized embedded deployment pipeline to minimize reliance on manual annotation and network bandwidth. A. System Overview The proposed architecture establishes a closed-loop ecosystem for continuous improvement of waste classification, as depicted in Fig. 1. The workflow begins with an Offline Training (left), where a scarcity of labeled real-world data was addressed by generating photorealistic synthetic images using generative AI. These images were used to train lightweight MobileNetV2 models, which are subsequently quantized to INT8 precision for deployment, as described next. The Edge Node (center) runs the trained models directly on an ESP32-S3-WROOM-1-N16R8 module, chosen for its AI vector instructions and 8 MB of PSRAM, and is connected to a wide-angle OV2640 camera. This hardware platform, initially described in [19], was designed for energy-autonomous operation in harsh outdoor environments. To save both energy and bandwidth, the device carries out inference locally and transmits results via NB-IoT only when a significant event occurs or the model’s confidence exceeds a defined threshold. The hardware consists of two boards housed within a single aluminium enclosure, selected for its durability, as shown in Fig. 2. The main board integrates the ESP32 module, camera, power management with solar charging, and the Quectel BC95-GR NB-IoT module. The secondary board includes a GPS module, temperature and humidity sensor, accelerometer, and an OLED display for debugging (secondary board is optional and supplementary to the main board when GPS is required or wake-up on motion is required etc.). A circular opening in the lid, covered with an acrylic window, serves as a viewing window for the OLED display during debugging. In deployment, this opening is covered with the solar panel and the OLED display is powered off. A rugged antenna housing is mounted on top of the enclosure, containing both GPS and NB-IoT antennas. Finally, the Cloud Infrastructure (right) serves as the central aggregation platform and user interface. Its role evolves through the system’s lifecycle. In the initial deployment phase, when the model is still being developed, the system is configured to transmit a higher volume of images to facilitate rapid validation of the model. Once field accuracy reaches a satisfactory threshold, the system shifts to a steady operational state. In this stage, full image transmission is triggered only by specific flags defined by the user or system logic, such as low inference confidence, detected anomalies (e.g., discontinuous fill level changes indicating unrecorded emptying), or material types that violate a specific waste contract. This selective IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 3 (a) (b) Fig. 2. Smart waste bin monitoring device hardware: (a) main PCB board showing the ESP32-S3-WROOM-1-N16R8 module, Quectel BC95-GR NB-IoT module, power management with solar charging, camera interface, and SD card slot; (b) 3D isometric cross-section view of the device in the protective housing. feedback loop ensures continuous adaptation while minimizing overhead. During adaptation, the model is fine-tuned on a mixed dataset comprising the new labeled real-world images and the original synthetic data to prevent retraining. B. Synthetic Data Generation To address the challenges associated with the acquisition and manual annotation of a large-scale, real-world dataset, we utilized a synthetically generated image dataset. A total of 285 photorealistic images were created using OpenAI’s GPT-Image-1 image generation model (accessed June 2025 using API), a generative artificial intelligence model. Generation used image-editing mode: an image of an empty skip bin served as the conditioning input, with edits constrained to the skip bin content (Fig. 3). Image of an empty skip bin was collected over number of days an interval of one hour to cover the most typical illumination conditions (Fig. 5). This approach facilitated precise control over the dataset composition through the use of structured, descriptive natural language prompts. The generation process was meticulously controlled to ensure a high degree of visual diversity and realism. Key parameters defined in the prompts included: •Skip bin characteristics: Attributes of the skip bin, such as “low-profile skip bin” or “weathered orange-brown steel,” were specified to maintain scene consistency. •Material type: Specific keywords such as “aerated concrete blocks,” “asphalt chunks,” “broken bricks,” and “gypsum debris” were used to render distinct waste categories (Fig. 6). •Scene composition: Critical variables including skip bin fill level (e.g., “60percent,” “100percent”) (see Fig. 4), camera perspective (“photographed from a low angle”), and lighting conditions (“daylight illumination”) were explicitly controlled. Since the deployment pipeline extracts the skip bin region of interest (ROI) prior to inference, the background environment surrounding the bin was irrelevant and held constant during generation. •Iterative Refinement: Prompts were iteratively adjusted to refine details like debris variation, texture, and the presence of contaminants, thereby enhancing the dataset’s realism. The generation process was automated using a Python script to ensure consistency and scalability. The core prompt template was defined as follows: “Fill the skip bin with [Material Type], filled to exactly [Fill %] of total capacity. Respect the perspective of the original image, especially when material is close to the camera. Keep the noise and quality of the original image.” Despite iterative prompt refinement, the generative model exhibited specific biases, particularly at high fill levels. The model tended to generate steep, centralized piles rather than filling the skip bin volume uniformly to the edges. Furthermore, it consistently avoided placing material in the immediate foreground (near the camera lens). This contrasts with realworld operational scenarios where debris often accumulates against the skip bin walls or directly occludes the camera view, representing a significant “hard case” for the synthetic-to-real transfer. To ensure consistency, the input images of empty skip bins were initially captured using a prototype device based on the ESP32-CAM AI-Thinker module, before a specific device was designed and manufactured. This device utilizes the same OV2640 sensor as the production unit, ensuring that the optical characteristics (lens distortion, sensor noise, color science) of the synthetic training data match the deployment environment. The final dataset was partitioned into a training and validation set of 221 images and a hold-out test set of 64 images. C. Data Preprocessing and Augmentation All images in the dataset were uniformly resized to an input resolution of 128×128 px. To enhance model generalization and mitigate the risk of overfitting on the synthetic data, a robust data augmentation pipeline was applied on-the-fly exclusively to the training set. The augmentation strategy, implemented using the TensorFlow Keras API, included: •Geometric transformations: Random horizontal flipping, random rotations within a range of ±20%, and random zooming up to 30%. IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 4 (a) (b) Fig. 3. Example of synthetic data generation: (a) Original real-world image of an empty skip bin captured by the prototype device. (b) Generatively augmented version simulating a skip bin filled with steel waste (Level 2), demonstrating the preservation of camera perspective and noise characteristics. (a) (b) (c) (d) Fig. 4. Synthetic generation of varying fill levels for the “Bricks” material category: (a) Level 0 (Empty), (b) Level 1 (25%), (c) Level 2 (50%), and (d) Level 3 (75–100%). The generative model maintains consistent lighting and texture while accurately rendering the increasing volume of debris. •Photometric transformations: Random adjustments to image contrast, brightness, and saturation. •Noise injection: Gaussian noise with a standard deviation of 0.25 was added to the images to improve model robustness. This comprehensive augmentation process synthetically expands the diversity of the training data, compelling the model to learn more invariant features. D. Model Architecture and Training To balance performance with computational efficiency, a transfer learning methodology was employed. Two independent classification models were developed: IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 5 (a) (b) (c) (d) (e) (f) Fig. 5. Variability in the input skip bin images showing different shadow patterns and illumination intensities: (a)-(f) represent images captured at different times of day. This natural variation ensures the model’s robustness to changing outdoor lighting conditions. 1) Material classifier: Identifies the waste type (e.g., Wood, Concrete, Plastic). 2) Fill level classifier: Categorizes the volume into four quartiles: Level 0 (0–25%, effectively empty), Level 1 (25–50%), Level 2 (50–75%), and Level 3 (75–100%). These models operate in tandem. A consistency check logic is applied during post-processing: a material classification of “Empty” implies a fill level of Level 0. Divergent predictions (e.g., Material=“Empty” with Level >0) are treated as lowconfidence anomalies, triggering the transmission of the full image for manual verification. The architecture for both models is based on a pre-trained MobileNetV2 with a channel width multiplier (alpha) of 0.35, leveraging weights from its training on the ImageNet dataset. A custom classification head was appended to the base model, consisting of a GlobalAveragePooling2D layer, a ‘Dropout‘ layer with a rate of 0.5 for regularization, and a final Dense layer with a softmax activation function. The training was conducted in two distinct stages: 1) Initial training: The convolutional base of the MobileNetV2 model was frozen, and only the newly added classification head was trained. This stage allows the model to learn task-specific features on the new head without altering the pre-trained weights. 2) Fine-tuning: Following the initial phase, the top 20 layers of the MobileNetV2 base were unfrozen. The entire model was then re-compiled with a significantly lower learning rate (1×10−5) and training was resumed. This fine-tuning stage allows the model to make small, specific adjustments to its pre-trained features to better adapt to the nuances of the waste dataset. Throughout both stages, the Adam optimizer and the categorical cross-entropy loss function were used. An EarlyStopping callback monitored the validation loss with a patience of 60 epochs, and a ReduceLROnPlateau callback dynamically adjusted the learning rate. E. Model Optimization for Edge Deployment Following training, the Keras models were optimized for on-device inference by converting them to the TensorFlow Lite (.tflite) format. Full integer quantization (INT8) was applied to reduce the model size and accelerate inference speed. This process converts all 32-bit floating-point weights and activations to 8-bit integers. A representative dataset, consisting of 100 images from the training set, was used to calibrate the quantization process, ensuring minimal loss of accuracy. II. RESULTS AND DISCUSSION The performance of the two trained models was evaluated based on their training dynamics and their accuracy on the hold-out test set. A. Training Dynamics The training history for both the material and fill level models demonstrates effective learning and successful regularization, as shown in Fig. 7 and Fig. 8. The loss curves for both training and validation sets show a consistent downward trend, while the accuracy curves steadily increase and then IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 6 (a) AAC (b) Asphalt (c) Bricks (d) Concrete (e) Glass (f) Gypsum (g) Misc (h) Polystyrene (i) Soil (j) Steel (k) Wood (l) Wool Fig. 6. Representative synthetic samples for 12 material categories. The generative pipeline produces distinct textures and debris geometries for each class. plateau at high values. The minimal gap between the training and validation curves indicates that the combination of strong data augmentation, a high dropout rate, and early stopping was effective in preventing significant overfitting. The fine-tuning stage, which begins after the initial epochs, shows a further slight improvement in both accuracy and loss, demonstrating the benefits of adapting the pre-trained layers. B. Model Performance on Test Data The models were evaluated on the hold-out test set of 64 images. The material classification model achieved an overall accuracy of 92.2%, while the fill level model achieved an accuracy of 95.3%. The detailed class-wise performance is presented in the confusion matrices in Fig. 9 and Fig. 10. For the material model (Fig. 9), the strong diagonal trend indicates a high rate of correct predictions across most classes. Many materials, including “empty,” “glass,” “gypsum,” and “steel,” were classified with 100% accuracy on the test set. The most significant confusion occurred between “bricks” and “concrete,” where 25% of “concrete” instances were misclassified as “bricks,” an understandable error given their visual similarity in texture and color. For the fill level model (Fig. 10), the performance is exceptionally strong. The model achieved 100% accuracy for Levels 0, 2, and 3. The only misclassification occurred with “Level 1,” IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 7 0.4 0.6 0.8 Accuracy Training accuracy Validation accuracy 0 50 100 150 200 Epoch 0.5 1.0 1.5 2.0 2.5 Loss Training loss Validation loss Fig. 7. Training and validation accuracy and loss curves for the material classification model over epochs. which was mistaken for “Level 2” in one instance, resulting in 86% accuracy for that specific class. This high level of precision validates the model’s suitability for determining the fill state of the skip bins. C. Real-World Validation To bridge the gap between synthetic training performance and real-world applicability, we conducted a pilot field test across two distinct skip bin types varying in geometry (0.5– 2.0 m depth, Fig. 11) and waste composition (wood, mixture of concrete and asphalt waste, and cobblestone). The device was mounted on the skip bin edge, tilted to capture the contents. To account for uneven filling patterns, measurements were taken from multiple positions along the skip bin’s edges. While ambient lighting was generally stable, the device’s positioning relative to bright or dark backgrounds introduced significant variability in auto-exposure and white balance, testing the model’s robustness to dynamic range shifts. The test conditions were challenging to evaluate the model’s behavior without retraining on completely different cases. The mixture of concrete and asphalt waste proved particularly difficult due to its unique source as laboratory sample waste, making it challenging to classify even by human inspection of the photographs. The heterogeneous nature of the waste material, where some areas contained more concrete, others more asphalt, and additional waste types, resulted in analysis results that varied with the camera position. This variability 0.3 0.4 0.5 0.6 0.7 Accuracy Training accuracy Validation accuracy 0 25 50 75 100 125 Epoch 0.75 1.00 1.25 1.50 1.75 Loss Training loss Validation loss Fig. 8. Training and validation accuracy and loss curves for the fill level classification model over epochs. is inherent to the validation approach, which evaluates the content from the camera’s perspective, just as a trained or experienced person would judge the skip bin contents based on the visible portion. In contrast, the wood waste represents typical construction and demolition waste. Notably, cobblestones were not included in the training dataset, presenting an out-of-distribution challenge. Additionally, the skip bin geometry differed significantly from those used in training. The depth of the test containers cannot be reliably inferred from photographs by human observers, and therefore the model is not expected to perform well on this dimension. It was not reasonably feasible to obtain identical containers to those used in training for the field test, though this was not considered a critical limitation. The actual deployment is performed on containers identical to those used in the initial training phase. To evaluate the system’s robustness in operational environments, field testing was conducted across a diverse range of waste skip bins. Fig. 12 presents the confusion matrix for the material classification from the collected inference events, verified against manual visual inspection of the transmitted images. The fill level classification results, which showed good performance with 89.4% accuracy, are presented in the confusion matrix shown in Fig. 13. Fig. 14 illustrates representative images captured by the deployed device, highlighting the variability in lighting conditions, material composition, and image resolution encountered in the field. The corresponding IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 8 AAC Asphalt Bricks Concrete Empty Glass Gypsum Misc Polystyrene Soil Steel Wood Wool Predicted label AAC Asphalt Bricks Concrete Empty Glass Gypsum Misc Polystyrene Soil Steel Wood Wool True label 50% 0% 0% 0% 0% 0% 0% 0% 50% 0% 0% 0% 0% 0% 40% 0% 50% 0% 10% 0% 0% 0% 0% 0% 0% 0% 0% 0% 94% 5% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 10% 90% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 100% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 100% 0% 0% 0% 0% 0% 0% 0% 66% 0% 0% 0% 0% 0% 0% 0% 33% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 100% 0% 0% 0% 0% 0% 25% 0% 0% 0% 25% 0% 25% 0% 25% 0% 0% 0% 0% 0% 0% 0% 50% 0% 0% 0% 0% 0% 50% 0% 0% 0% 0% 0% 0% 80% 20% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 16% 83% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 100% Accuracy: 68.9% Fig. 9. Normalized confusion matrix for the material classification model on the test set. The values represent the percentage of true instances for each class that were correctly or incorrectly classified. level 0 level 1 level 2 level 3 Predicted label level 0 level 1 level 2 level 3 True label 100% 0% 0% 0% 17% 17% 30% 34% 6% 6% 40% 46% 5% 0% 29% 64% Accuracy: 41.9% Fig. 10. Normalized confusion matrix for the fill level classification model on the test set. The model shows high accuracy across all four levels. quantitative assessment is presented in Table I, which details the raw inference confidence scores and predicted fill levels (a) (b) Fig. 11. Skip bins used for the verification. for the 800×600px resolution samples. Several limitations were identified during field validation that provide insights for future improvements. Wood content was generally misclassified as either empty or glass, likely due to the material’s shape—large format sheets differ significantly from the timber-like shapes used during training. This highlights the importance of including diverse material geometries in the training dataset to improve generalization. Additionally, completely dark images failed to be recognized, as such conditions were not represented in the training data. IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. X, MONTH 2025 9 TABLE I RAW INFERENCE CONFIDENCE SCORES FOR EACH CLASSIFIED PHOTO. LETTER LABELS ((A), (B), (C), ETC.) CORRESPOND TO THE IMAGES SHOWN IN FIG. 14. AAC Asphalt Bricks Concrete Empty Glass Gypsum Misc Polyst. Soil Steel Wood Wool Fill (–) (a) 45.3 9.8 23 0.8 2.0 4.7 – 3.9 7.4 0.4 – – 2.7 Level 2 (b) 4.3 1.2 30 – – 3.5 – 53.5 5.9 1.2 – – 0.4 Level 2 (c) 39.8 0.4 16.8 – – 0.4 0.8 11.3 29.7 – – 0.4 0.4 Level 2 (d) 48.8 0.4 0.8 – – 0.4 – 5.1 44.5 0.4 – – – Level 2 (e) 0.8 0.4 – 2 93.4 1.6 – – – 0.2 – – 1.6 Level 0 (f) 6.2 0.8 – 0.8 28.9 28.9 2.3 – 3.9 0.5 3.5 0.4 23.8 Level 0 (g) 1.2 2.3 – 0.4 32 62.5 – – – – – – 1.6 Level 0 (h) – 0.4 14.8 83.6 0.4 0.4 – – – 0.4 – – – Level 2 (i) 1.2 17.6 13.3 23.4 34.4 3.1 – – – – – – 7.0 Level 0 (j) – 0.4 0.4 – – – – 98 0.4 0.4 0.4 – – Level 2 (k) 9.8 49.6 33.6 1.6 – 0.8 – 1.6 0.4 0.6 – 1.2 0.8 Level 2 (l) 20.3 10.5 15.2 – 8.6 0.4 1.2 13.7 6.2 0.5 3.1 20.3 – Level 1 (m) 0.4 3.9 42.6 3.1 0.4 13.3 – 0.4 0.4 0.3 – – 35.2 Level 3 (n) 0.4 3.1 34.3 0.4 60.9 – – – – 0.9 – – – Level 0 (o) – 18 – 0.4 29.2 52 – – – – – – 0.4 Level 0 (p) 32 7.3 – 2 0.4 2 13.7 5.9 7.8 – 1.6 3.1 24.2 Level 0 (q) 0.8 0.4 2 0.8 93 2.3 – – 0.4 – 0.3 – – Level 0 (r) 0.8 2 0.8 0.8 80 10.5 0.4 2.7 – – 0.8 0.8 0.4 Level 0 (s) 3.1 11.7 1.6 0.8 72.3 7.8 0.4 1.2 0.4 0.3 – – 0.4 Level 0 (t) 0.8 60.2 1.2 0.8 12.8 8.2 7.4 3.1 – – 0.4 4.7 0.4 Level 0 AAC Asphalt Bricks Concrete Empty Misc Wood Predicted label AAC Asphalt Bricks Concrete Empty Misc Wood True label 100% 0% 0% 0% 0% 0% 0% 0% 100% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 25% 25% 25% 25% 0% 0% 0% 0% 0% 0% 0% 100% 0% 0% 37% 12% 0% 0% 0% 50% 0% 8% 0% 0% 0% 66% 0% 25% Accuracy: 66.7% Fig. 12. Normalized confusion matrix for the material classification model on real skip bins. The values represent the percentage of true instances for each class that were correctly or incorrectly classified. To mitigate this issue in deployment, the unit is configured to operate only during daytime hours, defined by wake and sleep times, thereby limiting the capture of dark images and ensuring more reliable inference results. Level 0 Level 1 Level 2 Level 3 Predicted label Level 0 Level 1 Level 2 Level 3 True label 80% 10% 10% 0% 0% 100% 0% 0% 0% 22% 77% 0% 0% 0% 0% 100% Accuracy: 89.4% Fig. 13. Normalized confusion matrix for the fill level classification model on real skip bins. The model achieved 89.4% accuracy, demonstrating robust performance across all four fill levels. D. Energy Efficiency Analysis To validate the power consumption characteristics, we profiled the device using a Nordic Power Profiler Kit II. Experiments were conducted indoors at a fixed location to minimize signal strength variability, isolating network protocol overhead (e.g., congestion, retransmissions) as the primary source of timing variance. To further isolate the energy cost of the vision and communication pipeline, the GPS module was physically removed. Each event contains a new network connection due