Designing and Developing an Advanced Drone-Based Pollution Surveillance System for River Waterways, Streams, and Canals Using Machine Learning Algorithms: Case Study in Shatt al-Arab, South East Iraq

Abstract

This study explores pollution detection and classification in the Shatt al-Arab River using advanced image processing techniques. Our proposed system integrates Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) algorithms. The Shatt al-Arab River in Basra, Iraq, faces increasing pollution from human activities, including oil spills, debris, and wastewater. We conducted extensive surveys of the river and its tributaries using a DJI Mavic drone, amassing over 1000 images to train machine learning models. The results indicate that RF excels with 94% accuracy for oil spills, 92% for wastewater, and 95% for debris. SVM also performs well, achieving 92%, 88%, and 94% accuracy for the respective pollutants. KNN, though insightful, lags with 85%, 89%, and 86% accuracy. Trained on this novel image dataset, these models show promising accuracy in detecting various pollution types from drone footage.

1. Introduction

River pollution poses severe ecological threats globally [1,2,3], with inadequate scientific monitoring, especially in data-scarce areas like Iraq’s Shatt al-Arab waterway. This diverse habitat faces dangers from wastewater, oil spills, and debris, causing water quality crises with significant health and environmental impacts, as documented in [4,5]. Despite local vulnerabilities, advanced vigilance is lacking. This study employs machine learning for the aerial identification of pollutants in the Shatt al-Arab River, utilizing drone surveys to automate multi-class pollution detection. The models, trained on an extensive image dataset collected by DJI Mavic Air 2 and Mini 3 4k drones, facilitate the timely detection of oil spills, debris, and sewage. The selected algorithms enable comparisons with previous studies and demonstrate the proposed model’s ability for real-time monitoring and pollution detection, addressing the longstanding challenge of environmental data scarcity in Iraq. This pioneering work introduces the first machine learning implementation for the pollution monitoring of Iraqi rivers, surpassing recent publications [6,7,8] with a simple, robust system achieving outstanding multi-class accuracy outcomes. In summary, this research pioneers drone-based water quality monitoring using machine learning in Iraq, advancing scientific understanding and supporting evidence-based decision-making for the protection of the imperiled river ecosystem, with the goal of transitioning models toward operational deployment and broader adoption.

In recent years, river pollution has become a critical issue, particularly in Iraq, with the Shatt al-Arab River facing challenges from industrial waste, sewage discharges, ship oil spills, and trash buildup. This pollution has significantly compromised water quality, posing threats to ecosystems, aquatic life, and public health. Reports from the Ministry of Environment in 2009 reveal that most Iraqi governorates do not meet national drinking water standards and WHO guidelines, with bacteriological contamination averaging 16%, ranging from 2.5% to 30% [9].

To address this crisis, developing an automated pollution detection system through machine learning is crucial. By utilizing advanced image classification algorithms trained on remote sensing and live feed data, the system can identify and monitor oil spills, trash, and wastewater in real-time [10,11,12]. Deployment along the Shatt al-Arab River would enable swift responses to pollution incidents, preventing major damage. This innovation has the potential to revolutionize river conservation efforts in Iraq, restoring the historic waterways.

By implementing such a system, proactive measures can be taken to reduce pollutants entering rivers, as water discharge releases harmful contaminants with the potential to cause diseases and adverse effects. Urgent and coordinated efforts are globally essential to combat various forms of pollution threatening rivers and to safeguard the environment, human health, and aquatic life [13].

The paper contributes to the field by addressing pollution detection in the Shatt al-Arab River, utilizing advanced image processing techniques integrated with machine learning algorithms. Leveraging drone technology, the study collects over 1000 images to train models to detect oil spills, wastewater, and debris. The results showcase high accuracy rates, with Random Forest achieving 94% for oil spills, 92% for wastewater, and 95% for debris. Support Vector Machines and K-Nearest Neighbors also perform well, albeit with slightly lower accuracy. The findings prove the proposed system’s potential as an efficient tool for monitoring pollution in the river, offering valuable insights for environmental management and conservation efforts in the region.

The paper is organized into several sections, each addressing different aspects of the research process. It begins with the Materials and Methods section (Section 2), which provides an overview of the study area and sampling design (Section 2.1), followed by a review of related work (Section 2.2). The Methodology section (Section 2.3) details the approach taken, including data collection (Section 2.3.1) and preprocessing (Section 2.3.2). Subsequently, Detecting River Pollution through Machine Learning (Section 2.4) delves into the core of the research, covering aspects such as the system interface (Section 2.4.1), feature extraction (Section 2.4.2), model selection for multi-pollutant detection (Section 2.4.3), model training (Section 2.4.4), testing (Section 2.4.5), deployment (Section 2.4.6), and continuous improvement (Section 2.4.7). The Results section (Section 3) presents the findings related to the image dataset (Section 3.1) and a performance evaluation of the models (Section 3.2), including evaluation metrics (Section 3.2.1) and results (Section 3.2.2). Following this, the Discussion section (Section 4) provides an analysis and interpretation of the results, while the Conclusions section (Section 5) summarizes the key findings and implications of the study.

2. Materials and Methods

2.1. Study Area and Sampling Design

In this study, an extensive survey was conducted of the Shatt al-Arab River and its tributaries in Basra, Iraq. Illustrating an area spanning approximately a 300 km distance, Figure 1 shows the twelve different canals and rivers that were covered. To obtain a detailed understanding of the region of the survey stations, aerial survey flights were carried out using two DJI series drones from November 2022 to May 2023 (see Appendix A). The data collected from these flights consist of approximately 1000 images and videos, allowing for a detailed examination of the riverbanks and the presence of various pollutants.

The twelve canals and rivers were chosen based on factors such as geographic distribution, accessibility, variety of environmental conditions, and known pollution hotspots (see Figure 2 for the pollution density distribution in Basra canals), along with the importance for ecosystem and human activities. This selection aimed to provide a representative sample of the pollution scenario across the Shatt al-Arab River system, allowing for a comprehensive understanding of pollution sources and their impacts on the region.

In the context of Iraq’s climate, particularly during periods of sunny, clear skies and partly cloudy conditions, the UAV data collection was effective. Iraq experiences a dry and arid climate, especially in regions like Basra where the Shatt al-Arab River is located. During periods of sunny weather with clear skies, the visibility for UAV flights was optimal, allowing for high-resolution imaging and video capture. These conditions provided excellent lighting for capturing detailed images of the survey area, enhancing the quality and accuracy of the collected data. Partly cloudy conditions, depending on the extent of cloud cover, introduce some variability in the lighting conditions. However, during data collection, the visibility remained adequate, with partly cloudy conditions being still suitable for UAV operations as there was no rain. Overall, sunny and clear weather conditions are preferred for UAV data collection in Iraq, ensuring optimal visibility and high-quality data capture for the study of the Shatt al-Arab River.

The findings of the survey revealed the presence of wastewater, oil spills, debris, and other forms of pollution in the Shatt al-Arab River and its tributaries. These pollutants pose a significant threat to the ecosystem of the river and its surrounding areas. This information is crucial for the development of effective environmental management strategies in Basra, as it provides an excellent understanding of the pollution sources and other environmental challenges faced by the region. These observations can lay the groundwork for additional research and planning aimed at preserving and restoring the ecological health of the river and its surrounding areas.

2.2. Related Work

In paper [6], the authors introduce a sophisticated model integrating computer vision and deep learning for image classification purposes. This innovative approach employs mounted cameras on bridges overlooking rivers to gather visual data, harnesses crowdsourcing for the manual labeling of plastic materials, and develops a two-stage deep neural network structure. Utilizing Faster R-CNN and Inception V2 models to identify and classify plastic objects, the research method evaluates model performance through cross-validation and compares the outcomes with human visual counts. In the study, 1272 JPEG images captured from bridges over rivers at 5 locations in Jakarta, Indonesia, were used to train a segmentation model, and then 14,968 images of floating objects labeled as plastic or not by crowdsourcing were used to train a classification model.

The base plastic detection model achieved 59.4% precision on the test set without optimization. Applying data augmentation, an Adam optimizer, and learning rate decay improved the precision to 68.7%. Extrapolation indicated further gains of up to 73% precision with more training data. When tested on new locations, the precision dropped to 20–54% without retraining. Retraining with just 50 new examples improved the precision to 42% on a challenging new site. On 82 samples, the model counts showed reasonable agreement with human visual counts (coefficient of determination R2 = 0.43). The model detected 35% more plastics, on average, compared to human counters.

A group of researchers [14] investigated the correlation between the total nitrogen (TN) concentrations extracted from drone-captured hyperspectral imagery of plants and experimentally measured TN concentrations in the Ebinur Lake Oasis. The research utilized four machine learning models, namely, PLS, RF, ELM, and GP, to assess Pearson’s Correlation Coefficients and identify significant wavelengths in the red-edge regions. By implementing a fractional-order derivative model, the correlation between the variables improved, achieving an absolute PCC value of 0.63. Then, 45 water samples were collected from sites along the Aqikesu River in Ebinur Lake, China, and used to measure the total nitrogen (TN) concentration in the lab to develop a ground truth dataset. Hyperspectral imagery was captured over sampling sites using a drone-mounted Headwall Nano-Hyperspectral camera (400–1000 nm range). Thirty samples were selected via the Kennard-Stone algorithm for model training. The remaining 15 samples were reserved for model testing. Then, 200 datasets of 45 samples were generated via sampling with replacement from the original 45 used to train the ML models. A fractional-order derivative (FOD) was used to enhance spectral features. The FOD increased the correlation between plant spectra and TN concentration. In addition, 23 sensitive wavelengths were identified using FOD-processed spectra. Four ML models were trained on bootstrapped samples to estimate TN: Random Forest (RF), partial least squares (PLS), Gaussian Process (GP), and extreme learning machines (ELM). The GP model performed best individually, with a coefficient of determination R2 of 0.88 and mean square error (MSE) of 0.52 mg/L on calibration data. All individual models showed high variability in terms of accuracy metrics. The decision-level fusion (DLF) model achieved the highest accuracy (R2 = 0.99) by combining ML models. DLF reduced the uncertainty and minimized the biases of individual models. The validation results also showed high accuracy for DLF (R2 = 0.84, RMSE = 0.5 mg/L).

Another group of researchers [13] presented a methodology combining image segmentation, classification, and optical flow to detect and track marine oil spills using aerial imagery. Despite its limited scope, the study demonstrated the potential of deep learning techniques for environmental monitoring and response. Enriching the training data and broadening pollution types could bolster adoption. The study used a limited amount of data, with only 435 images being used for training and testing. Out of the total data, 300 images were used for pollution detection and 135 images were used for non-pollution detection. Image segmentation with a fully convolutional network was used to identify pollution regions. Segmented regions were classified as pollution/non-pollution using SVM. The optical flow between the frames was calculated to estimate the movement direction of the polluted regions. Aerial images of a real ship oil spill off the coast of Taiwan were used for testing. Segmentation plus SVM achieved 99.95% accuracy in detecting oil-polluted regions. The optical flow visualization demonstrated the capability to track the spread of oil spills. Limitations include a small training dataset size and limited pollution types. The study showed potential for automated marine monitoring and oil spill response using AI and drones.

Paper [15] proposed a novel technique to quantify riverine plastic debris based on unmanned aerial vehicles (UAVs), specifically focusing on Malaysia’s Klang River. Visual counting measurements were taken during fieldwork between 29 April and 4 May 2019, from the Jalan Tengku Kalana bridge, while UAV-borne measurements were collected downstream at 700 m. Comparing these results with traditional methods revealed that UAV-based monitoring offers an alternative to the monitoring of riverine plastic transport, particularly in inaccessible and remote locations.

A related study [7] outlined an innovative methodology that employs drones and deep learning techniques to detect and map marine litter in coastal regions. The authors emphasized the severity of the marine litter issue, particularly plastics, and underscored the limitations of existing monitoring methods in terms of coverage and efficiency. The researchers enlisted commercial off-the-shelf drones operated by citizens to gather aerial images of beaches. These images were then segmented into tiles and annotated by volunteers with classifications such as “litter” or “no litter” utilizing an online platform, ultimately generating training data. Subsequently, five deep convolutional neural network architectures were trained on these annotated tiles to autonomously identify marine litter. A total of 1975 ultra-resolution images (5472 × 3648 pixels with an aspect ratio of 3:2) were captured using the DJI Phantom 4 Pro v2 quadcopter. The Phantom, with a 20-megapixel camera and a mechanical shutter, was mounted to a three-axis gimbal. The raw images were divided into smaller tiles, each measuring 512 × 512 pixels, and were annotated using the Zooniverse tool. The authors divided the tiles into two classes, namely, litter and not litter. The proposed citizen science protocol using commercial drones and crowdsourced image labeling enabled the efficient collection of a large training dataset of over 30,000 labeled image tiles. The drone metadata extraction process allowed the georeferencing and spatial analysis of results. The VGG19 architecture achieved the best performance with 77.6% accuracy on the held-out beach test set for litter detection. The density maps visualized the concentration and distribution patterns of litter reasonably well compared to manual surveys. The deep learning model litter detection results aligned with manual visual classifications with an R2 of 0.97 between density maps.

A similar study [8] put forth a pragmatic approach for the automatic mapping of plastic waste in rivers by combining the use of UAVs and deep learning models. This method has the potential to benefit a diverse array of stakeholders involved in monitoring and mitigating plastic pollution in waterways. By leveraging UAVs and deep learning models, the precise detection of plastic litter in rivers—including rural areas and transboundary rivers—becomes practical and efficient. The pre-trained YOLOv5s model was identified as particularly effective for detecting plastic in riverine UAV imagery, striking an optimal balance between accuracy and computational complexity. Moreover, transfer learning techniques significantly enhanced the model performance, thereby improving plastic detection with minimal additional training required. The insights gleaned from this research contribute to the ongoing global efforts to decrease plastic waste and safeguard innumerable aquatic ecosystems. The key data used in this paper include aerial images captured by a DJI Phantom 4 drone from surveys of two rivers—the Houay Mak Hiao river in Laos and the Khlong Nueng canal in Thailand. The images covered 2 m × 2 m patches of the rivers. Here, 500 aerial image tiles of 256 × 256 pixels were extracted and manually annotated with bounding boxes around plastic objects for each river location. These formed the training and test datasets. Metadata like GPS coordinates and altitude were extracted from the raw drone images to georeference the tiles. The Houay dataset had 592 annotated plastic objects and the Thailand dataset had 796 objects. The tiles were split 70% for training and 30% for validation/testing for the deep learning models. The paper evaluated the performance of different deep learning models for plastic detection in rivers using UAV sensor data. YOLOv5s was found to be the most effective model, with a high mean average precision (mAP) of 0.81 with no transfer learning based on the Houay Mak Hiao dataset. Transfer learning was applied to improve the performance of the models. Pre-trained YOLOv4 with transfer learning achieved the highest accuracy, namely, a 3.0% increase in mAP to 0.83. YOLOv3 also showed itself to be significant for transfer learning, with an increase in mAP from 0.59 to 0.81. The pre-trained YOLOv5s model used the Houay Mak Hiao dataset provided the best tradeoff between accuracy and computational complexity.

2.3. Methodology

This research develops an advanced pollution patrol program developed for the Shatt al-Arab waterway. Detecting pollution in rivers using machine learning algorithms is a challenging but important task for environmental monitoring. To create a pipeline for this work, we prepared multiple steps, namely, data collection, preprocessing, feature extraction, model training, and evaluation, as shown in Figure 3.

A DJI Mavic Air 2 drone with high-performance cameras collects photographic and video data depicting the river’s ecosystem. Following data collection, careful preprocessing filters noise and highlights key features. Next, benchmarking assesses three machine learning models—RF, SVM, and K-NN. By quantifying accuracy, precision, and recall during extensive training and validation, the top-performing algorithm is identified. This algorithm will enable the patrol program, providing conservationists with valuable artificial intelligence tools. Applying these AI-driven insights, new policies can restore the waterway’s ecological health.

2.3.1. Data Collection

The acquisition of data for this investigation, which is depicted as “dataset” in Figure 3, was accomplished through the deployment of unmanned aerial vehicles (UAVs), which procured high-resolution imagery pertaining to pollution in the Shatt al-Arab locale. Two models of commercially available UAVs were employed—the DJI Mavic Air 2, equipped with a 1/2” CMOS 48 MP camera yielding 4 K/60 fps video, and the DJI Mini 3, furnished with a 1/1.3” 12 MP camera enabling 4 K video recording. Flight times near 30 min facilitated broad coverage. Survey locations included Shatt al-Arab stretches frequently impacted by oil spills, sewage drainage points, and litter-accumulated shorelines near populated areas. Scheduling flights across various times (early morning, afternoon) and seasons (summer, winter) captured diversity, with early morning flights detecting overnight discharges. Imported JPEG and MP4 files were decomposed into frames.

In the interval between November 2022 and April 2023, examinations were conducted over critical regions of the Shatt al-Arab River as well as its adjacent tributaries and waterways. During these six months, over 300 km of UAV flight trajectories were executed to image the survey zones. One such trajectory is illustrated in Figure 4. The high-resolution visual data encompassed diverse forms of pollution such as river debris, oil spills, refuse, and wastewater. Cumulatively, more than 18 h of high-definition aerial footage was obtained, encompassing over 1000 images and videos encapsulating the primary classifications of contamination in situ. This extensive dataset presents varied real-world manifestations and serves as a crucial basis for devising and verifying machine learning algorithms aimed at detecting and categorizing pollution incidents.

2.3.2. Data Preprocessing

The collected drone data underwent rigorous preprocessing using the Python programming language and the scikit-learn library [16]. The frames from the captured MP4 files were resized to 450 × 450 pixels, retaining key visual components, enabling faster processing. Multiple data augmentation techniques were applied, including horizontal flipping, rotations, and brightness adjustments to boost model generalization.

Model Development: 80% of the data (801 samples) was used to train the RF, SVM, and KNN models, with 20% (200 samples) held out for testing. Hyperparameters like number of RF trees, SVM kernel type, and KNN k-value were tuned through grid search and cross-validation. Model selection was based on the highest recall, to reduce false negatives on the test set.

Integration and Testing: The chosen model was integrated into our system overlaying the real-time drone camera feed with the pollution detection results. This system underwent field testing for oil spill and trash identification.

2.4. Detecting River Pollution through Machine Learning

The heart of the pollution detection system lies in the integration of machine learning models. These models provide the intelligence to analyze images, videos, and data from drone flights, allowing the application to identify pollution events accurately.

2.4.1. Interface of the System

The interface serves as the front end, connecting users to the models’ predictions and insights. Figure 5 displays the interface that allows the user to view the captured video. The processing of the video starts after clicking on the Start button. For every second, the pollution presence is detected, as can be seen in the terminal. The linked drone’s position is determined to identify its inclination while flying.

2.4.2. Feature Extraction

Color-based features are often used in oil spills, which often create distinctive color patterns on the water surface, including dark black and iridescent rainbow sheens [17]. However, varying illumination, perspective, and occlusions can impede robust feature extraction from imagery. Oil’s optically dense properties engender divergent visual signatures compared to surrounding waters, stemming from the differential interaction with solar radiation, particularly intense absorption in the 8000–14,000 nm spectrum. The resulting variations in long-wave infrared coloration facilitate the identification of surface slicks. To leverage these discernible color cues, we extract color-based features from aerial images captured by drones surveying the Shatt al-Arab River [18]. Specifically, we convert RGB images into HSV color space, which decouples hue, saturation, and value information [19]. HSV enables targeted feature engineering, a technique widely utilized in image analysis [20]. Our feature vector incorporates the mean and standard deviation of H, S, and V components across images, capturing color distribution statistics. Additionally, color histogram descriptors are extracted to represent the distribution of hues. By learning from such color-based features, the machine learning model can identify oil spill color patterns to automatically detect their presence in drone-captured imagery.

Innovatively utilizing the HSV color space for feature extraction in identifying oil spills poses several challenges in distinguishing pollution from natural water features. These challenges include varying illumination, perspective distortions, occlusions, and the complex optical properties of oil slicks. There were multiple challenges to be addressed. Firstly, illumination changes across images lead to inconsistent color representations. To mitigate this, statistical measures such as mean and standard deviation of hue, saturation, and value components were calculated across images. These measures provide a more robust representation of color distribution, reducing the impact of varying illumination. Secondly, some drone-captured images exhibit perspective distortions due to the angle of observation. This affects the appearance of oil slicks and natural water features. However, by leveraging statistical features extracted from the HSV color space, the model learns to recognize consistent color patterns associated with oil spills despite perspective distortions. Another challenge was represented by objects obstructing the view of the water surface, hindering the identification of pollution. However, oil slicks have distinct optical properties that differ from the surrounding water, allowing for their detection even in partially occluded areas. Our machine learning model was trained on color-based features and it learned to differentiate between oil slicks and other objects, improving the detection accuracy. Oil slicks represented another challenge because they exhibit unique optical properties, such as intense absorption in specific spectral ranges. By focusing on these spectral characteristics, the model identifies surface slicks based on their distinct color patterns in the long-wave infrared spectrum, facilitating their detection in drone-captured imagery.

By addressing these challenges and leveraging discernible color cues from the HSV color space, the approach enhances the detection of oil spill color patterns, enabling the automated identification of pollution in aerial imagery of the Shatt al-Arab River.

2.4.3. Model Selection—Algorithm Selection for Multi-Pollutant Detection

Given the diversity of pollution types and real-world complexities of the expansive Shatt al-Arab drone dataset, algorithm selection was critical for optimizing multi-class contaminant identification. Support Vector Machines (SVMs), K-Nearest Neighbors (k-NNs), and Random Forests (RFs) were chosen for the task of detecting river pollution in the study of the Shatt al-Arab River in Basra, Iraq, for multiple reasons. SVM excels at recognizing intricate patterns in complex data [21]. This is particularly crucial when identifying different types of river pollution, such as wastewater, oil spills, and debris.

Another study [22] exhibited diverse and non-linear patterns. k-NN was used because it is highly suitable for spatial analysis. Pollution detection in a river often depends on the geographical proximity of data points. k-NN can effectively identify areas with similar pollution characteristics based on their spatial relationships [23]. RF was employed because it is known for its ability to handle large and noisy datasets. In this study, with over 1001 images and videos from aerial surveys, RF can provide reliable results by combining multiple decision trees to reduce overfitting and handle data noise effectively.

All three models were chosen because they offer scalability and efficiency, crucial for processing a substantial amount of visual data from the aerial survey flights. They can efficiently handle large datasets, making them a practical choice. k-NN and RF were preferred for interpretability and visualization. They provide insights into the most influential features for pollution detection, helping researchers understand the causes of pollution and develop effective management strategies. The flexibility of parameter tuning for these models allows customization to the specific pollution detection task. Adjusting parameters in SVM, k-NN, and RF enables fine-tuning for optimal performance tailored to the dataset and research goals [24].

In summary, SVM was chosen for its pattern recognition capabilities, k-NN for spatial analysis, and RF for its handling of large, noisy datasets. All three models were selected due to their scalability and efficiency in processing extensive visual data, while k-NN and RF were favored for their interpretability and customization through parameter tuning. These choices collectively provide a robust framework for identifying and addressing river pollution in the Shatt al-Arab River and its tributaries.

2.4.4. Model Training

A Python script controls the DJI Mavic drone to capture images while flying over designated waypoints, representing areas of interest such as oil spills, river debris, and wastewater. It uses the DJI Mobile SDK to manage the drone’s flight and image capture. The images are associated with specific labels for oil spills, river debris, and wastewater.

Relevant features are extracted from images using three different machine learning models: SVM, k-NN, and RF. These models are trained on the labeled images, and their accuracy in classifying the images is evaluated, providing insights into the performance of the models for this image classification task. The dataset is split into 70% training, 10% validation, and 20% testing sets. The code is modular, allowing for the customization of image preprocessing and the definition of categories for classification.

The three machine learning models included parameter tuning. For the SVM model, hyperparameter tuning involved exploring options for the choice of kernel, penalty parameter C, and kernel coefficient gamma. This tuning process aimed to find the optimal combination of these parameters that maximizes the model’s performance in classifying images. With regard to the kernel type, in case of the linear kernel, it assumes a linear decision boundary between classes. Another type, the polynomial kernel, introduces polynomial features and allows for non-linear decision boundaries. We explored degrees 2 and 3. The third option to be investigated was the radial basis function kernel that considers the similarity between data points in high-dimensional space. The penalty parameter C was varied. For C = 0.1, it allowed a softer margin, allowing some misclassifications. In case of C = 1, it balanced between maximizing the margin and minimizing misclassifications. For C = 10, a stricter penalty was imposed on misclassifications, leading to a narrower margin. The kernel coefficient gamma was varied, namely, 0.1, 1, 10, as these values determine the influence of a single training example, with higher values giving more weight to nearby points. The best option chosen based on the hyperparameter tuning process for the SVM model was the radial basis function as a kernel type because it provided the best performance in capturing the complex relationships present in the image data. The penalty parameter C was set to 1 to obtain a balance between maximizing the margin and minimizing misclassifications. The kernel coefficient gamma was set to 0.1, as it effectively captures the similarity between data points in high-dimensional space without overfitting. These options were determined to yield the highest accuracy and generalization performance. As such, they represent the optimal configuration for the SVM model in classifying images captured by the drone over the Shatt al-Arab River region.

For the k-NN model, hyperparameter tuning involved selecting the optimal number of nearest neighbors (k) to optimize the model’s performance in classifying images. This tuning process involved testing various values of k and selecting the one that yielded the best results on the validation set. The tested numbers of nearest neighbors were 3, 5, 7, and 10. For k = 3, the classification was based on the three nearest neighbors, which captured local patterns in the data effectively. In the case of k = 5, it allowed for a broader consideration of neighboring points, capturing more global patterns in the data. Another tested value was k = 7, which led to smoother decision boundaries and reduced the impact of noise in the data. The last tested value, k = 10, provided an even broader perspective by considering the classification based on ten nearest neighbors. Upon completion of the tuning process, the selected value of k = 5, determined to yield the best performance, represented the optimal configuration for the k-NN model in classifying the images captured by our system.

For the RF model, hyperparameter tuning involves selecting the optimal number of trees in the forest, the maximum depth of the trees, and the minimum number of samples required to split a node. The number of trees in the forest determines the robustness and stability of the model. More trees lead to better performance, but also increase the computational cost. The testing values included a range from 100 to 500 trees with increments of 100. This range allowed us to evaluate the impact of different numbers of trees on the model performance. The maximum depth of trees controls the complexity of individual trees in the forest. Deeper trees capture more complex relationships in the data, but also lead to overfitting. The testing values included a range from 10 to 50 with increments of 10. This range covered a spectrum of tree depths, from relatively shallow to deeper trees, allowing for the exploration of different levels of complexity. The minimum number of samples required to split a node determines the minimum number of samples required to split an internal node during the tree-building process. Increasing this value helps to prevent overfitting by ensuring that nodes with insufficient samples are not split. The testing values included a range from 5 to 50 with increments of 5 that reflected a gradual increase in the minimum number of samples required to split a node. This range allowed us to test the impact of different levels of node splitting criteria on the model generalization. After thorough hyperparameter tuning, the final decision for the RF model’s parameters consisted of a forest of 300 trees that yielded the best balance between accuracy and computational efficiency. After assessing the model’s performance with varying tree depths, it was found that setting the maximum depth of the trees to 20 resulted in optimal generalization performance without overfitting. Through experimentation with different values for the minimum number of samples required to split a node, it was concluded that setting this parameter to 10 achieved the most robust and stable model performance. These parameter values were chosen based on their ability to maximize the model accuracy and generalization while minimizing overfitting. They represented the optimal configuration for the RF model in classifying the images captured by the drone over the Shatt al-Arab River region.

2.4.5. Model Testing

The model performance was rigorously assessed for SVM, k-NN, and RF using the validation dataset. A K-fold cross-validation approach [25] enables robust and reliable algorithm evaluation during training, as all samples are used in the training, validation, and testing, which leads to a more generalizable and stable model. Evaluation metrics including accuracy, precision, recall, F1-score, and area under ROC curve (AUC) gauged each model’s proficiency. This multi-faceted evaluation methodology facilitated a thorough understanding of the pollution detection performance of the models across diverse scenarios.

The ultimate validation of the models, including SVM, k-NN, and RF, was conducted using the test dataset to simulate real-world scenarios and assess their practical utility. The dataset was split into 70% training images (701), 10% validation images (100), and 20% testing images (200) to provide an overview of the dataset distribution. This threefold dataset division ensured that each model was rigorously tested on an independent dataset, mimicking real-world conditions, and enhancing the reliability of the findings. Evaluation metrics, including accuracy, precision, recall, F1-score, and the area under the ROC curve (AUC), were then systematically applied to gauge the models’ performance in practical pollution detection scenarios in the Shatt al-Arab River.

2.4.6. Deployment

To make the research findings accessible and actionable, the trained models, including SVM, k-NN, and RF, were integrated into a real-time monitoring system, as shown in Figure 6, which is tailored for river pollution detection. These models were developed as part of the current research using a dataset of videos and images related to river pollution. These models have been fine-tuned and tested to classify or detect specific types of river pollution, such as oil spills, debris, and wastewater. Figure 6 is a graphical representation illustrating the structure and operation of the tested real-time monitoring system. It shows how data are collected, processed, and fed into the trained models for pollution detection. This visualization helps in understanding the overall system’s design.

2.4.7. Continuous Improvement

Recognizing the dynamic nature of river ecosystems and pollution patterns, a strategy for continuous improvement was instated. This encompassed a periodic model (e.g., 5 min) based on a demo that updates using fresh data and the integration of supplementary data sources such as remote sensing data and meteorological information to bolster the model accuracy.

3. Results

The effectiveness of any pollution detection system relies on the diversity and quality of input images. In our research, we curated a dataset of river images and videos capturing wide-ranging pollution scenarios. These images provide the critical foundation on which machine learning models are built and validated.

3.1. Image Dataset

Our dataset encompasses a wide spectrum of environmental conditions and pollution scenarios, including instances of:

• Oil spills: Images capturing the telltale signs of oil spills in water bodies, characterized by distinctive color patterns and textures.
• Wastewater contamination: Scenes depicting the presence of discolored and contaminated water resulting from wastewater discharge.
• River debris: Visual evidence of floating debris and garbage in river segments, endangering aquatic ecosystems.

To provide a visual understanding of the input images used in our research, we present two sample images—one depicting pollution and another one without pollution—in Figure 6.

Figure 6 illustrates a pollution scenario in which an oil spill and wastewater are visible on the water’s surface, along with garbage. The characteristic dark and iridescent patterns of the oil spill and wastewater are evident in the image, highlighting the presence of pollution. In contrast, Figure 7 represents a non-polluting scenario, showcasing a pristine river segment with clear water and no visible signs of pollution. This serves as a reference for clean and healthy river environments.

3.2. Model Performance Evaluation

The success of the machine learning-based approach for pollution detection in river ecosystems hinges on its ability to accurately classify and detect instances of pollution. This section outlines the performance evaluation of the algorithms: SVM, k-NN, and RF. These models were rigorously evaluated using the validation dataset to measure their proficiency in detecting pollution events.

3.2.1. Evaluation Metrics

The assessment of algorithm performance in this study employed various evaluation metrics, as outlined by [26,27,28,29]. These metrics include accuracy, precision, recall, F1-score, and the area under the ROC curve. Accuracy is the ratio of correctly classified instances to the total instances in the validation dataset. Precision represents the proportion of correctly predicted positive observations relative to the total predicted positive observations, indicating how accurately positive predictions are made. Recall is the ratio of correctly predicted positive observations to all actual positive observations, measuring the model’s ability to capture all positive instances. F1-score, the harmonic mean of precision and recall, offers a balanced evaluation of a model’s overall performance. The area under the ROC curve (AUC) serves as a metric quantifying the model’s capability to distinguish between positive and negative classes, with higher AUC values indicative of superior performance.

3.2.2. Performance Results

Table 1. Performance results of SVM, k-NN, and Random Forest for the three pollution classes.

Model	Class	Accuracy	Precision	Recall	F1-Score	AUC
SVM	Oil spill	0.92	0.91	0.93	0.92	0.96
	Wastewater	0.88	0.86	0.89	0.87	0.90
	Debris	0.94	0.95	0.93	0.94	0.97
k-NN	Oil spill	0.85	0.84	0.87	0.85	0.88
	Wastewater	0.89	0.88	0.90	0.89	0.92
	Debris	0.86	0.87	0.85	0.86	0.89
RF	Oil spill	0.94	0.93	0.95	0.94	0.97
	Wastewater	0.92	0.91	0.93	0.92	0.95
	Debris	0.95	0.94	0.96	0.95	0.98

summarizes the performance of the SVM, K-NN, and RF models based on the evaluation metrics.

With regard to the overall model performance, SVM exhibits consistent and commendable performance across all classes, with high accuracy, precision, recall, and AUC values. k-NN demonstrates competitive results, maintaining a balance between precision and recall. Random Forest stands out as the top-performing model, consistently achieving the highest scores in accuracy, precision, recall, F1-score, and AUC for all three pollution classes.

SVM performs consistently, with solid precision and recall. k-NN is moderately competent, but trails behind Random Forest and SVM. Lastly, for debris, Random Forest excels in detection, achieving top scores across metrics. SVM maintains consistent performance with high precision and recall. k-NN delivers comparable results, though marginally behind SVM and Random Forest.

The choice of the most suitable model depends on the specific requirements of the pollution detection application. In terms of trade-offs, decision-makers may need to balance factors such as precision, recall, and computational complexity based on the priorities of the environmental monitoring system. The environmental impact implications for efficient resource allocation are to understand that the relative strength of each model allows for a more efficient allocation of resources, potentially optimizing the environmental conservation efforts.

A detailed breakdown facilitates targeted actions against specific types of pollution, enabling more effective and precise interventions. Random Forest emerges as a robust and reliable choice across all pollution classes, making it well suited for practical deployment in real-world scenarios. In the case of SVM and k-NN, while slightly behind Random Forest, these models still deliver commendable results, and their use may be justified based on specific application requirements. In summary, the study underscores the efficacy of machine learning models, with Random Forest exhibiting excellent overall performance. Considerations for class-specific strengths should guide the selection of models based on the specific pollution challenges faced in river ecosystems.

In our quest to detect pollution in river ecosystems using machine learning algorithms, we rigorously evaluated the performance of three distinct models: SVM, k-NN, and RF. To gain deeper insights into their performance, we analyzed the confusion matrices, which provided a granular view of how each model performed in classifying pollution occurrences. The confusion matrix provides a deep interpretation of a model’s performance; for example, if the data were unbalanced, namely, skewed towards one class, the model accuracy will provide misleading information and the proposed model may fail to obtain the same level of accuracy when testing with a new, unseen dataset. Additionally, some models may be robust in predicting one class while performing poorly in detecting another class or classes. Therefore, reporting the sensitivity and specificity values are beneficial in this work. Sensitivity (true positive rate) and specificity (true negative rate) are critical metrics for evaluating model performance, particularly in the context of pollution detection. Sensitivity measures the model’s ability to correctly identify positive instances, while specificity gauges its ability to correctly identify negative instances. By analyzing the confusion matrices, we can identify areas where sensitivity and specificity need improvement. For instance, if sensitivity is high and specificity is lower, this indicates a need to reduce false positives.

The confusion matrices for SVM, k-NN, and Random Forest are presented in

Table 2. Confusion matrices for SVM, k-NN, and Random Forest.

Model	Class	True Negative (TN)	False Positive (FP)	False Negative (FN)	True Positive (TP)
SVM	Oil spill	7	5	58	280
	Wastewater	1	8	33	308
	Debris	1	4	13	332
k-NN	Oil spill	15	5	80	250
	Wastewater	7	2	58	283
	Debris	4	5	68	273
RF	Oil spill	1	2	47	300
	Wastewater	2	3	40	305
	Debris	1	1	14	334

.

In the context of oil spill detection, the SVM model demonstrates high accuracy, particularly for true positives. However, it is not without flaws, showcasing instances of misclassification involving both false positives and false negatives. For wastewater and debris, the model exhibits commendable performance, with a strong emphasis on true positives and minimal false positives in the case of debris.

The k-NN model, while slightly trailing behind SVM in terms of oil spill detection accuracy, maintains a balanced performance with a mix of true positives and false positives. In wastewater detection, it demonstrates good overall performance, and in debris identification, it mirrors the performance of SVM, exhibiting high values for true positives and low values for false positives.

The Random Forest model emerges as a standout performer, showcasing high accuracy and robust performance across all metrics for oil spill detection. Its consistency in delivering good performance, especially for true positives, and outstanding accuracy with minimal misclassifications in debris identification positions it as a formidable choice.

In an overarching evaluation, the Random Forest model consistently outperforms both SVM and k-NN across all pollution classes. Its strength lies in its ability to excel in correctly identifying pollution instances (high true positive rate) while simultaneously minimizing false identifications (low false positive rate).

To calculate sensitivity and specificity from the confusion matrix, we used the following formulas:

Sensitivity (True Positive Rate) = TP/(TP + FN)

(1)

Specificity (True Negative Rate) = TN/(TN + FP)

(2)

Based on the confusion matrix from Table 2, the sensitivity and specificity for each model was calculated (

Table 3. Sensitivity and specificity values for SVM, k-NN, and Random Forest.

Model	Class	Sensitivity	Specificity
SVM	Wastewater	0.903	0.111
SVM	Debris	0.962	0.200
k-NN	Debris	0.801	0.444
RF	Oil spill	0.864	0.333

), and it can be observed that SVM is best used for wastewater and debris detection; k-NN manages debris detection well, but less than SVM, while RF is good for oil spill detection.

For all three models, SVM, k-NN, and RF, the sensitivity values ranged from approximately 0.801 to 0.962 across different pollution classes. This suggests that the models perform well in correctly identifying instances of pollution when they are present. The specificity values vary across models and pollution classes, ranging from approximately 0.111 to 0.444. Lower specificity values show that the models may exhibit a higher rate of false positives, incorrectly identifying non-polluted areas as polluted. Balancing sensitivity and specificity is crucial in pollution detection. Optimizing both metrics ensures that the models accurately identify pollution occurrences while minimizing false alarms, thereby facilitating effective environmental monitoring and management efforts. Model refinement efforts, as future work, will focus on adjusting the classification thresholds, feature engineering, and exploring ensemble methods to optimize sensitivity and specificity simultaneously, enhancing the overall performance and reliability of the pollution detection models.

The findings of this comparative analysis validate the utility of machine learning algorithms, particularly the Random Forest model, for robust and reliable pollution detection in river ecosystems. The high accuracy, precision, recall, and AUC values indicate the potential practical deployment of these models in real-time monitoring systems, contributing significantly to environmental conservation efforts.

4. Discussion

The study identified an accurate algorithm for pollution detection in the Shatt al-Arab region based on real-world data. In addition, the findings have significant implications for environmental conservation efforts, offering a technological solution to support the preservation of the Shatt al-Arab ecosystem.

The evaluation of SVM, k-NN, and RF models underscores the importance of a tailored approach in pollution detection. While each model exhibits strengths, the Random Forest model emerges as the top performer, demonstrating accuracy and reliability. These results not only pave the way for practical applications in environmental monitoring, but also contribute to the ongoing preservation and enhancement of the health of aquatic ecosystems. The continuous refinement and optimization of these machine learning models are anticipated to further enhance their effectiveness in real-world scenarios, marking a significant step forward in environmental stewardship.

In summary, the Random Forest model clearly emerges as the top technique for this river pollution classification task based on the empirical evaluation across multiple performance metrics. The automated image-based river pollution monitoring system should be built using the Random Forest algorithm to ensure robust and reliable detection. The SVM also holds potential as a strong backup technique. Further research should explore combining these supervised learning models with deep neural networks to potentially improve generalizability.

Comparing the results from [6] to our experimental evaluation, there are some notable similarities and differences. The previous paper obtained 68.7% precision for plastic detection using an optimized deep learning model, while our RF model achieved a higher accuracy of 94% on the river pollution image dataset. The paper model’s precision dropped significantly when tested on new locations. Our other models achieved perfect accuracy. Our controlled dataset contributed to higher accuracy.

In paper [14], drone hyperspectral imagery was used to estimate the TN concentration in water bodies, achieving an R2 of 0.99 with their model. Our system uses 4k resolution images from Shatt al-Arab River to detect multiple pollution types, attaining 92–94% accuracy with SVM and Random Forest models. While the previous paper focuses specifically on nitrogen, our system detects various pollutants like plastics, oil, and waste. Our system obtained comparable accuracy levels—SVM with 92% accuracy and 0.96 AUC, and Random Forest achieving 94%. The previous paper uses machine learning on image features, while we employ deep learning on raw images for end-to-end classification. Our system’s strengths are as follows: it uses real-world images from Shatt al-Arab rather than drone data; it detects diverse pollution types beyond just plastics; the deep learning leverages are for end-to-end learning directly from images; it is a computationally efficient deployable system rather than using offline analysis; it achieves high accuracy without requiring hyperspectral data. In addition, deep learning algorithms benefit from a large dataset, as the weights in each layer are better tuned with more data points (training data), expectedly leading to higher accuracy compared to traditional machine learning models; thus, in the pilot study of the current paper and with the size of data that we used, utilizing SVM, KNN, and RF was more appropriate than using a deep learning model [30].

In paper [13], SVM classification was used to achieve segmentation accuracy ranging from 92.5 to 99.5% for oil detection and 85.7% for trash detection. The dataset was limited and unbalanced, so it is unclear how the system will perform when testing with a larger dataset. Our system attained 92–94% accuracy for pollution classification using SVM, RF, and KNN with a large amount of data. Our system is applicable to different pollution types—debris, oil, sewage water, etc. It leverages real visual data from Shatt al-Arab and provides versatile real-time pollution detection using readily available visual data and efficient algorithms.

The novelty of this research lies in the use of unmanned aerial vehicles for extensive aerial surveys of river ecosystems. By employing drones equipped with cameras, the study captures high-resolution imagery of the Shatt al-Arab River and its tributaries, enabling a detailed examination of pollution sources and environmental conditions. The application of machine learning algorithms, including SVM, KNN, and RF, for pollution detection in river ecosystems is novel because we observed three pollutants, namely, oil spills, wastewater, and debris. These algorithms analyze the collected imagery data to automatically identify and classify pollutants, providing valuable insights for environmental management strategies. The use of color-based features extracted from images using the HSV color space represents an innovative approach to pollution detection. By decoupling hue, saturation, and value information, this technique enables targeted feature engineering, enhancing the models’ ability to detect distinctive color patterns associated with different types of pollutants on the water surface. The current study conducted an evaluation of model performance using confusion matrices and sensitivity and specificity metrics, providing a granular view of each model’s classification capabilities. This comprehensive assessment ensures robustness and reliability in pollution detection, facilitating informed decision-making for environmental management.

The high accuracy of our tested SVM, RF, and KNN models holds significant implications for real-world applications in environmental management. For instance, the ability to accurately identify and classify pollutants in aerial imagery can facilitate the early detection of environmental hazards, allowing for timely intervention and mitigation measures. This can help protect the ecosystem of the Shatt al-Arab River and its surrounding areas, safeguarding the health and well-being of local communities reliant on the river for various activities such as fishing, agriculture, and recreation. Furthermore, these models can support decision-making processes for environmental agencies and policymakers by providing valuable insights into pollution hotspots, trends over time, and potential sources of contamination. By leveraging the predictive capabilities of these models, stakeholders can implement targeted interventions to reduce pollution levels, restore ecosystem health, and ensure the sustainable management of the river ecosystem.

Despite the promising performance of the models, several limitations were encountered during development and evaluation. One notable challenge was the availability and quality of the training data, which can introduce biases or inconsistencies in model predictions. Additionally, the models struggled to generalize to unseen environmental conditions or respond effectively to changes in pollutant characteristics or distribution patterns. Moreover, the computational resources required for the real-time application of these models in large-scale monitoring operations could pose practical challenges, particularly in resource-constrained settings. Addressing these limitations will require ongoing efforts as future work to refine the model algorithms, improve data collection methods, and enhance collaboration among stakeholders to ensure the models’ relevance and reliability in real-world applications.

Addressing potential challenges in scaling or adapting the pollution detection system to other regions is crucial for understanding its broader applicability. Different regions may exhibit varying environmental characteristics, such as water turbidity, vegetation cover, and land use patterns. These variations can impact the performance of pollution detection models trained on data from a specific region. Adapting the system to new regions requires careful consideration of these environmental factors and may necessitate the retraining or fine-tuning of the models to account for local conditions. Regulatory frameworks and socioeconomic conditions vary across regions, influencing pollution sources, reporting mechanisms, and community engagement practices. Adapting the system to new regions requires collaboration with local stakeholders, including government agencies, environmental organizations, and community groups, to ensure alignment with regulatory requirements and to address community needs and concerns. Cultural attitudes towards environmental monitoring and data sharing may differ across regions, impacting the acceptance and adoption of the pollution detection system. Ethical considerations, such as privacy rights and data ownership, must be addressed when deploying monitoring systems in diverse socio-cultural contexts. By discussing these potential challenges, stakeholders can gain a more rounded view of the system’s applicability and develop strategies to overcome barriers to scaling and adaptation in diverse geographic and socioeconomic settings.

5. Conclusions

In our study, we developed a pollution detection system using drone imagery and machine learning techniques. Our Random Forest model attained a substantially higher accuracy of 94% compared to the 77.6% accuracy of the top VGG19 model in [7]. Our precision, recall, F1 scores, and AUC were consistently higher across all tested models, including SVM and KNN. This indicates an overall better performance. This was likely due to better flight planning and conditions. We classified more than three classes, compared to their one category of litter. In conclusion, our methodology provides higher detection accuracy on drone imagery. This is likely due to our optimized data collection process, better dataset, multi-classification task, and more consistent manual labeling. Our high-performance metrics demonstrate the strengths of our approach over this initial work on river pollution detection using drones and deep learning.

Future work will involve integrating the optimized algorithm into an operational pollution monitoring system. Moreover, it would be good to expand the study to other regions facing similar environmental challenges that can enhance the broader applicability of the developed UAV program and algorithms.

In conclusion, the presented methodology outlines the systematic approach to developing a UAV-based pollution monitoring program and evaluating machine learning algorithms for pollution detection. The research aims to provide a valuable tool for environmental conservationists and policymakers in the Shatt al-Arab region and beyond.