Hardware calibration and testing

The Raspberry Pi-based multi-sensor data logging system was designed following a structured approach that included hardware integration, software development, calibration and system validation. Figure 1 shows the system architecture and design. The hardware configuration comprised a Raspberry Pi 3 microcomputer connected to a DS18B20 temperature sensor, a microphone and a camera module.

Figure 1. System architecture and data flow diagram.

A 5 V power supply unit powered the system, and the Raspberry Pi’s General Purpose Input/output (GPIO) pins were configured to interface directly with the sensors (Vujović and Maksimović, Reference Vujović and Maksimović2014). The system’s architecture enabled real-time data acquisition, processing and storage, while also supporting remote access via a RealVNC server linked to an internet-connected remote viewer Vijay et al., Reference Vijay, Abishek, Sabarish and Krishnan2023). During assembly, all hardware components were mounted securely and configured to support synchronized environmental data collection. Python scripts were developed to automate sensor data reading, logging and storage in structured formats, such as Comma Separated Values (CSV). The RealVNC server setup allowed users to monitor data remotely and control the system in real time. Calibration of the DS18B20 temperature sensor was carried out under controlled laboratory conditions using a mercury-in-glass thermometer for reference. Figure 2 shows the data calibration of the temperature sensor. The sensor was mounted externally on the experimental water container to minimize thermal interference from the Raspberry Pi’s internal heat and to allow for rapid heat dissipation.

Figure 2. Data calibration for the temperature sensor.

The camera module was calibrated through adjustments in focus, white balance and exposure settings to ensure optimal image clarity across varying lighting conditions. The microphone was tuned using the Linux-based “arecord” tool to enhance underwater sound capture. To validate the system’s waterproofing and sensor functionality, submersion tests were conducted in two stages: initially in a 35 cm deep water tank, and subsequently in a 250-l drum at 1-m depth. These trials verified sensor durability, enclosure integrity and stable data acquisition under aquatic conditions. Observations from these tests informed iterative improvements in the waterproof casing design to ensure consistent sensor performance and data reliability during field deployments.

AI model training and validation

AI algorithms were developed to enable automated species identification in Lagos Lagoon by analyzing both visual and acoustic data. Hydrophones were deployed at multiple depths and locations within the lagoon to capture underwater soundscapes for acoustic monitoring (Lillis et al., Reference Lillis, Caruso, Mooney, Llopiz, Bohnenstiehl and Eggleston2018). Convolutional neural networks (CNNs) and recurrent neural networks were considered for classifying the captured audio signals based on known vocalizations of aquatic species. These ML models were trained using labeled datasets containing species-specific acoustic patterns (Datta et al., Reference Datta, Maharaj, Prabhu, Bhowmik, Marino, Akbari, Rupavatharam, Sujeetha, Anantrao, Poduvattil, Kumar and Kleczkowski2021) and were continuously refined through retraining and validation to improve classification accuracy and reduce false positives. In parallel, a deep learning model based on CNNs was implemented for image-based aquatic species identification. The visual dataset comprised 31 distinct species, which were partitioned into training (8,791 images), validation (2,751 images) and testing (1,760 images) subsets. To enhance model robustness and generalization, data augmentation techniques, such as rescaling, shearing, zooming and horizontal flipping, were applied. The CNN architecture featured four convolutional layers with increasing filter sizes (from 32 to 256), followed by batch normalization, max pooling and dropout layers to prevent overfitting. A global average pooling layer was incorporated to reduce spatial dimensions, feeding into fully connected dense layers with L2 regularization. The final output layer used softmax activation to support multi-class classification. The model was compiled with the Adam optimizer (learning rate = 0.001) and categorical cross-entropy as the loss function. Model performance was tracked using accuracy and loss metrics across epochs. Early stopping criteria and dynamic learning rate adjustments were applied to optimize training. The final model demonstrated the capability to recognize species based on visual and acoustic data, contributing significantly to biodiversity monitoring in dynamic aquatic environments.

System evaluation

The performance, reliability and practical deployment of the low-cost multi-sensor system were carried out in both laboratory and field settings. The evaluation focused on key operational parameters, including sensor drift, biofouling resistance, power efficiency and data transmission reliability over time (Chan et al., Reference Chan, Schillereff, Baas, Chadwick, Main, Mulligan and Thompson2021). Comparative assessments with commercial-grade sensors were also planned to benchmark the accuracy and feasibility of the low-cost alternatives. Additionally, stakeholder engagement with local researchers and environmental agencies was initiated to evaluate the system’s scalability and utility for long-term biodiversity monitoring. System integration tests were conducted to ensure interoperability among the temperature sensor, camera, microphone and Raspberry Pi-based data logging unit. Custom Python scripts were executed simultaneously across all components to verify real-time concurrent data acquisition. The system was networked and accessed remotely via a RealVNC server, demonstrating its capability for remote monitoring and control. Power efficiency was tested using a 22.2 kWh backup battery, which provided ~6 h of continuous operation. To evaluate environmental durability, staged submersion trials were conducted using a 250-l water drum at a depth of ~1 m. These trials confirmed the waterproof enclosure’s integrity and the system’s ability to record temperature, acoustic and visual data in submerged conditions. While real-time Wi-Fi data transmission failed beyond a 10 cm submersion depth due to signal attenuation, the system continued logging data locally, allowing for full retrieval upon resurfacing. The results of these evaluations validated the robustness and operational viability of the low-cost sensor system, confirming its suitability for extended deployment in dynamic aquatic environments, such as Lagos Lagoon.

Field deployment

The low-cost sensor system in (Figure 3a,b) was deployed on December 17–18, 2024, at Lagos Lagoon (6°30′27″N, 3°24′14″E) to collect real-time data on habitat conditions and aquatic biodiversity. Figure 4 illustrates the deployment process, highlighting both daytime and nighttime operations. The sensors were strategically installed at a fixed location to ensure consistent and reliable data collection. A stable power supply and reliable connectivity are required to facilitate continuous monitoring and data transmission. Before field deployment, the system underwent laboratory testing, where it successfully collected temperature data, images and audio recordings. To enhance durability in aquatic environments, the sensors were enclosed in a waterproof casing and initially tested in a controlled water tank (35 cm depth). During early trials, minor water ingress was observed, requiring three design iterations to achieve an effective waterproof seal. The moment the enclosure was fully sealed, further testing was conducted at a depth of 1 m using a 250 L water drum. The system maintained full functionality under these conditions, confirming its waterproofing effectiveness. However, a major limitation observed was in real-time data transmission. The Wi-Fi signal was lost after ~10 cm of submersion, consistent with known challenges in underwater wireless sensor networks, where signal attenuation and propagation delays hinder effective communication. Despite this limitation, all recorded data were stored locally within the Raspberry Pi, allowing for data retrieval upon resurfacing.

Figure 3. (a) Hardware system setup. (b) Complete system set up.

Figure 4. (a) Configuration of the system before deployment. (b)Deployment of sensors at the Lagoon water during the daytime. (c) Deployment of sensors in the Lagoon water during the night.

Temperature measurement

The temperature variations recorded by the deployed sensor are presented in the graph, with temperature (°C) on the vertical axis and time on the horizontal axis. The temperature data were taken at 2-s intervals and converted to minutes’ average. The data indicate a gradual increase in temperature, peaking at ~31.36 °C around 08:05 PM, followed by a slight decline, as shown in Figure 5. These fluctuations may be attributed to environmental factors, such as solar heating, water movement or sensor positioning. This study highlights the transformative potential of AI-powered multi-sensor systems in environmental research. The device demonstrated reliability in temperature monitoring and organism detection, reinforcing its applicability for aquatic ecosystem studies. However, further refinements are necessary to enhance accuracy and adaptability across diverse aquatic conditions (Abdelwahab, Reference Abdelwahab2023).

Figure 5. Temperature recorded by the monitoring system.

Acoustic measurement

The acoustic measurements were collected at Lagos Lagoon, capturing underwater sound activity over time, as shown in Figure 6. The top graph is a spectrogram, which visualizes sound intensity across different frequencies. The horizontal axis represents time, while the vertical axis represents frequency (Hz). The color intensity denotes amplitude, where darker regions indicate lower sound levels, and brighter regions reflect stronger signals. The spectrogram reveals varying frequency distributions, likely influenced by aquatic life, boat traffic or environmental noise. The lower graph, the magnitude spectrum, highlights dominant frequency components, showing that most recorded acoustic energy is concentrated below 500 Hz. This suggests that the primary sound sources in the area covered by the lagoon operate within the low-frequency range, a characteristic commonly associated with both biological and anthropogenic underwater noise. These findings corroborate with Kershenbaum et al. (Reference Kershenbaum, Akçay, Babu-Saheer, Barnhill, Best, Cauzinille and Dunn2025), who reported that automatic detection of biological sounds enables faster and more efficient analysis of large acoustic datasets, a crucial advancement for processing long-duration field recordings. Furthermore, methods for localizing sound sources help minimize discrepancies in recordings, leading to more accurate identification and quantification of individual sound producers.

Figure 6. Sound wave spectrum recorded (top: amplitude spectrum, middle: frequency spectrum, and bottom: magnitude spectrum).

Visual data collection

Figure 7 presents sample images of organisms captured during the pilot field test. The device successfully documented aquatic vegetation, including water hyacinths. However, due to the brackish nature and high turbidity of the Lagos Lagoon, the images appeared dark, limiting visibility and detail. These findings emphasize the need for improved lighting or image enhancement techniques to enhance data quality in future deployments. Datta et al. (Reference Datta, Maharaj, Prabhu, Bhowmik, Marino, Akbari, Rupavatharam, Sujeetha, Anantrao, Poduvattil, Kumar and Kleczkowski2021) reported that the use of custom-built multispectral cameras, along with advanced data augmentation techniques, such as spectral transformations and synthetic data generation, can improve dataset diversity and account for complex environmental conditions. Integrating similar approaches could enhance image clarity and increase the accuracy of visual data analysis in challenging aquatic environments.

Figure 7. Dry stalk of water hyacinth (Eichhornia crassipes).

AI machine learning model deployment

In Table 1, we report the first 20 epochs of the training results for the CNN model with key performance metrics. The training and validation accuracy gradually increased but remained relatively low due to the complexity of the dataset. As shown in Figure 8, both training and validation loss exhibited a downward trend, suggesting the model was learning, albeit at a slow rate. The learning rate remained constant at 0.001 throughout the training. Figure 9 demonstrates that the ML model effectively captures images in line with the trend of the observed dataset, indicating a reasonable ability to generalize to unseen data. However, the noticeable deviations suggest that further optimization is required, such as deploying a more powerful graphics processing unit (GPU). This may include increasing the model’s complexity by adding additional layers or filters, applying data augmentation to improve the diversity of the training data and fine-tuning hyperparameters, such as the learning rate and dropout rate. After training, the AI model was deployed on a Raspberry Pi within a multi-sensor data logging system designed for species detection in aquatic ecosystems. If recognition failed, the image was stored without an event log. This process continued iteratively until the monitoring campaign concluded, ensuring continuous data collection and analysis in real-world conditions.

Table 1. Training performance metrics over the last five epochs

Figure 8. Loss and validation loss for the image recognition machine learning model.

Figure 9. Machine learning model accuracy for image recognition.