Quantification of Suspended Sediment Concentration Using Laboratory Experimental Data and Machine Learning Model

Abstract

Monitoring sediment concentration in water bodies is crucial for assessing water quality, ecosystems, and environmental health. However, physical sampling and sensor-based approaches are labor-intensive and unsuitable for large-scale, continuous monitoring. This study employs machine learning models to estimate suspended sediment concentration using images captured in natural light, named RGB, and near-infrared (NIR) conditions. A controlled dataset of approximately 1300 images with SSC values ranging from 1000 mg/L to 150,000 mg/L was developed, incorporating temperature, time of image capture, and solar irradiance as additional features. Random forest regression and gradient boosting regression were trained on mean RGB values, red reflectance, time of captured, and temperature for natural light images, achieving up to 72.96% accuracy within a 30% relative error. In contrast, NIR images leveraged gray-level co-occurrence matrix texture features and temperature, reaching 83.08% accuracy. Comparative analysis showed that ensemble models outperformed deep learning models like Convolutional Neural Networks and Multi-Layer Perceptrons, which struggled with high-dimensional feature extraction. These findings suggest that using machine learning models and RGB and NIR imagery offers a scalable, non-invasive, and cost-effective way of sediment monitoring in support of water quality assessment and environmental management.

1. Introduction

Monitoring sediment concentration in water bodies is essential for managing water quality, protecting aquatic ecosystems, and ensuring public safety [1]. Sediment concentration influences turbidity, nutrient distribution, and pollutant transport, impacting water quality and ecosystem health [2,3]. Traditional sediment measurements must collect samples and conduct laboratory analysis, which are labor-intensive and costly for large-scale or real-time monitoring [4]. Conventional approaches to estimating suspended sediment concentration include direct water sampling with laboratory analysis [4], deployment of in-stream turbidity sensors [2], and physically based hydrological models such as SWAT or HEC-RAS. Although widely used, these techniques still demand substantial field effort, entail high costs, provide limited temporal resolution, and require extensive site-specific calibration. Moreover, process-based models require detailed input parameters and can struggle in data-scarce or highly dynamic settings. These constraints have accelerated interest in data-driven alternatives, especially machine-learning frameworks—which can capture complex, nonlinear relationships between imagery, environmental variables, and sediment transport processes. The need for efficient and scalable measurement solutions has driven researchers to explore non-intrusive remote sensing technologies and machine learning models as viable alternatives for estimating sediment concentration under diverse environmental conditions.

Recent advances in imaging and remote sensing have enabled RGB and infrared (IR) imagery for environmental monitoring [5], providing non-intrusive, high-resolution data of suspended sediment concentration (SSC). In the study [6], UAVs equipped with RGB cameras were used to correlate red-band reflectance with SSC, allowing for high-spatial-resolution mapping of sediment distribution. Such studies underscore the utility of spectral data in sediment analysis, offering insights into SSC patterns that can guide feature selection for machine learning models. Additionally, ref. [2] highlights advancements in optical and acoustic sensors for SSC measurements, highlighting the importance of high-frequency data collection in dynamically changing environments. Together, these studies motivate the exploration of machine learning models for estimating sediment concentration using remotely sensed spectral images in the site measurements.

Machine learning models, notably ensemble- and tree-based approaches, have shown promise in environmental data prediction due to their robustness and ability to manage complex, nonlinear relationships. For example, ref. [7] demonstrated that machine learning classifiers, particularly Support Vector Machines (SVM) could accurately assess microbial contamination by incorporating sediment-related features alongside traditional water quality indicators. Similarly, ref. [8] emphasized the potential of random forest (RF) for water quality prediction, using accessible parameters like temperature, pH, and conductivity to predict nitrogen and salinity levels accurately. These studies underscore the importance of feature selection in environmental monitoring tasks, highlighting the role of specific features, such as RGB means, red reflectance, time of image capture, and temperature, in optimizing model accuracy for sediment concentration. In addition, SVM has been applied in various water quality classification tasks and is effective at capturing nonlinear relationships in high-dimensional spaces [9], though it can be sensitive to hyperparameter tuning [7] and noise. The K-Nearest Neighbors (KNN) algorithm, while simple and interpretable, has also been used in environmental modeling [9], but it tends to underperform when applied to high-dimensional datasets or noisy inputs.

Additionally, one study [10] found that gradient boosting models, particularly LightGBM and XGBoost, achieved exceptional accuracy in water quality prediction by capturing complex feature interactions and variability in river basin data. Another example, ref. [11], demonstrated that Convolutional Neural Networks (CNNs) can be highly effective for particle concentration classification. While CNNs typically require large datasets, their study addressed this challenge through transfer learning and data augmentation.

Among the machine learning approaches explored in prior research, RF has gained prominence in water quality modeling owing to its robustness, interpretability, and ability to rank feature importance. Studies such as [12] highlight RF’s ability to handle nonlinear relationships and provide reliable predictions for water quality parameters. Similarly, ref. [8] show RF’s adaptability in predicting diverse water quality indicators, from nitrogen to salinity levels, based on various environmental features. Given its proven accuracy in water quality predictions, these studies validate RF’s applicability for sediment concentration modeling. To complement RF’s strengths, we use gradient boosting regression (GBR) as a parallel model, as it has shown strong performance in water quality prediction tasks [8,10].

Gradient boosting models (GBMs) effectively handle noisy data and generate interpretable feature importance rankings, making them well-suited for sediment concentration prediction tasks. For instance, ref. [13] demonstrates that LightGBM and XGBoost achieve high accuracy in microbial water quality prediction and provide insights into key influencing features such as lake turbidity when combined with SHAP analysis. GBMs, particularly when combined with dimensionality reduction or feature selection techniques, have shown strong adaptability in high-dimensional water quality prediction tasks [10,13]. Similarly, ref. [14] used gradient boosting with PCA to classify water quality, showing adaptability to high-dimensional data.

Building on the overview discussed, SSC is traditionally measured through direct sampling, sensor-based approaches, or remote methods. Direct sampling involves collecting water samples using depth- or point-integrated samplers, filtering them in a laboratory, drying the retained sediment, and weighing it to calculate SSC (e.g., in mg/L or ppm) [15,16]. Optical sensors estimate SSC by analyzing water turbidity, which is linked to light backscatter from suspended particles, while acoustic sensors use sound backscatter; both methods require calibration using sediment samples [17]. Remote sensing offers a non-intrusive alternative, but few studies have explored how water reflectance in satellite and aerial images can be used to infer SSC [5,6,18]. For example, ref. [19] demonstrated that MODIS satellite imagery could be assimilated into a 3D hydrodynamic model to improve SSC predictions in tidal environments, highlighting the value of combining satellite data with numerical modeling. Image-based SSC estimates are influenced by site-specific conditions, including sediment properties (e.g., particle size, mineral content), water quality (e.g., conductivity, total dissolved solids), and other environmental factors [16]. These challenges have prompted the use of machine learning techniques that leverage remotely sensed imagery and environmental variables for more efficient SSC prediction.

Remote sensing is an increasingly valuable tool for measuring SSC, particularly through natural visible and infrared imagery. For instance, ref. [6] demonstrated the utility of red-band reflectance in correlating SSC with high spatial accuracy. Their findings underscore the value of spectral data for sediment analysis, which directly inspired this project’s emphasis on RGB mean values and red reflectance. Similarly, ref. [18] demonstrated that optical satellite sensors like Terra MODIS could be used to estimate sediment concentrations across a range of conditions, leveraging temporal changes in reflectance patterns during high-flow wet seasons. Ref. [5] extended this concept by developing a low-cost, close-range remote sensing system that uses consumer-grade DSLR cameras with multispectral capabilities to estimate SSC, demonstrating the feasibility of red-band and near-infrared reflectance for sediment monitoring. However, these technologies often require expensive equipment and site-specific calibrations, motivating the need for cost-effective, data-driven alternatives like machine learning.

Despite these advances, a comprehensive laboratory-developed dataset that combines natural light and IR imagery is unavailable for predicting sediment concentration. This study addresses this gap by evaluating RF, GBR, CNN, MLP, SVM, and KNN on a recently laboratory-developed dataset. The dataset includes sediment concentration measurements taken under controlled conditions, allowing for rigorous model training, calibration, and validation. This study also proposes a novel machine learning framework for predicting sediment concentration using features extracted from RGB, near-infrared (IR) images, and temperature- and time-based information. RFR and GBR models were trained on mean R, G, and B values, red reflectance, and temporal features for natural light images, achieving robust predictive accuracies for both RFR and GBR. Gray-level co-occurrence matrix (GLCM) texture features combined with temperature data allowed the models to capture unique sediment characteristics for infrared images. These findings highlight the potential of integrating natural and infrared light imagery using machine learning models for scalable, non-invasive sediment monitoring, offering a valuable tool for environmental management and water quality assessment. Given the constraints of our dataset, we selected RFR and GBR for their proven ability to achieve reliable accuracy even with limited data. These models are well-supported by prior studies, which highlight how effective feature selection and model tuning can capture the complexities of sediment-related imagery data.

The following sections outline the methodology, including data preprocessing, model development, feature selection, and evaluation metrics. A comprehensive analysis of model performance and feature importance is provided, focusing on the strengths of RFR and GBR for SSC prediction using both natural and NIR images. This work contributes a scalable, data-driven approach for SSC measurements, supporting the growing demand for efficient and real-time water quality management in dynamic aquatic environments.

2. Experimental Runs and Observation Data

2.1. Experimental Set-Up

The dataset used in this study was developed in a controlled laboratory environment to ensure precision and consistency in observations. We set up ten 120 L volume black-colored rectangular water containers outside in the hydraulics laboratory at the University of Arizona. At the beginning of each experiment, we filled up each container with about 80 L of clear water. Then, we weighed sediments according to the targeted concentrations in each container and then added different amounts of sediment to each container with target concentration from 1000 ppm to 150,000 ppm. Afterwards, because dry sediment from field sites has some organics, we filtered out these floating debris from the surface until the surface was completely clear.

Each container had a tag showing the case ID numbers, which indicated the time of measurements and whether the images were captured using RGB or NIR lights. To ensure comprehensive temporal representation, images were collected at four times of the day: 9:00–11:00 a.m., representing morning light; 11:00–1:00 p.m., representing nearly at noon; 3:00–5:00 p.m., representing afternoon light; and 6:00–8:00 p.m., representing after dawn (early evening light).

2.2. Image Collection

At each collection time, sediment in each container was stirred up using a mixer until nearly all sediment was suspended and none was left at the container bottom, and then water temperature and solar radiance were measured. Simultaneously, water with sediment was sampled from the surface, and multiple images were taken using natural light cameras. The water samples were weighed and recorded immediately. Then, we repeated the same procedure for all the containers. The measured SSCs correspond to each set of images. After completing the RGB images at a given time, we repeated the same procedure for taking the NIR images. NIR lighting has a wavelength of 850–940 nm, which is typical of NIR cameras. Approximately 1300 images were captured to represent sediment-laden flow conditions in a laboratory tank, with SSCs ranging from 1000 mg/L to 150,000 mg/L.

To calculate the SSC, we measured the volume of the sampling container (984 mL). The weight of the container filled with clear water was obtained before the experiments. Therefore, the SSC in parts per million (ppm) was calculated using the following equation:

$$C\left(\mathrm{p}\mathrm{p}\mathrm{m}\right)={\displaystyle \frac{(W_t-W_b-984\times 0.9915)\times 1000}{0.984\times (1-\frac{0.9915}{2.65})}}$$

(1)

where W_t is the total weight of water, sediment, and bottle in grams, W_b is the weight of the bottle in grams, the density of water is 0.9915 g/mL, and the specific weight of sediment is 2.65.

Each image was assigned a unique ID and tagged with metadata including SSC, temperature, solar radiance, and the time of measurements.

2.3. Image Processing

All images were preprocessed by cropping to remove noise or visual disturbances and to ensure consistent resolution across all images. Figure 1 contains four natural images and four near-infrared light images that are cropped with unique SSCs. Each image captured under natural light was processed to extract key color-based features. Specifically, the image was loaded as a three-channel RGB array using the skimage.io module from the Scikit-Image library in Python (https://scikit-image.org/), which is commonly used for reading and processing images in RGB format, and the mean intensity for each channel—red, green, and blue—was calculated by averaging all pixel values within that channel. These mean RGB values represent the overall color composition of the image and serve as core input features for the machine learning models. Additionally, a normalized red reflectance metric was computed by dividing the average red intensity by the sum of the average intensities across all three channels (R/(R + G + B)). While this differs from the calibrated red reflectance used in [6], the underlying principle remains consistent, as both capture red spectral dominance, which has been shown to correlate with suspended sediment concentration.

For near-infrared (NIR) images, preprocessing involved converting each image to grayscale using the skimage.io.imread (i.e., for gray = true) function. Each grayscale image was then augmented through geometric transformations such as horizontal and vertical flipping, as well as rotation (e.g., 90°, 180°). From both the original and augmented grayscale images, texture-based features were extracted using the gray-level co-occurrence matrix (GLCM) method. Specifically, GLCMs were computed at multiple pixel distances (e.g., 1, 5, and 10) and angles (0, 45°, 90°, and 135°), capturing spatial relationships between pixel intensities. Statistical texture descriptors—including contrast, homogeneity, energy, correlation, dissimilarity, and entropy—were calculated using the “graycoprops” function from the skimage feature module, where “graycoprops” computes statistical texture properties (like contrast, energy, homogeneity) from a gray-level co-occurrence matrix (GLCM) in an image. These features quantitatively represent the spatial texture and structure of suspended sediments in NIR images, providing critical inputs for the machine learning models. Time of image capture and water temperature were also included as additional features for each sample. This robust dataset is the foundation for the machine learning models evaluated in this study, enabling the prediction of sediment concentration across a diverse range of lighting conditions and sediment densities.

3. Methodology

3.1. Theoretical Basis

Clean water absorbs longer wavelengths of light, such as red, orange, and yellow, while scattering shorter wavelengths like blue and green [20]. As a result, pure water appears blue. However, when mixed with sediment, its color changes depending on the size of the particles, the inherent sediment color, and the presence of contaminants or organic matter attached to them.

Larger particles (such as sand) primarily cause Mie scattering, which tends to scatter light across all visible wavelengths with little wavelength dependence, though it strongly favors forward scattering [21]. This makes water containing sand appear warmer-toned, with yellow or brown hues. In contrast, finer particles (such as silt and clay) exhibit Rayleigh-like scattering, preferentially scattering shorter wavelengths (blue and green). However, despite this tendency, organic matter and associated fine sediments can increase absorption in the blue and green spectral regions, contributing to the brownish color of water. This effect is further influenced by the particle composition and refractive index, as noted by [22].

Furthermore, many suspended sediments exhibit yellow, brown, or reddish hues due to iron oxides and other mineral constituents. These minerals enhance reflectance in the red and near-infrared regions of the spectrum, contributing to the warm-toned appearance of sediment-laden waters. This spectral behavior is well-documented in optical studies of natural waters and is particularly evident in environments with high concentrations of inorganic particulate matter [20,23]. As a result, water typically appears increasingly yellow or brown as the concentration of suspended sediment rises, influenced by particle size, organic matter content, and iron content in the sediment. Iron content determines sediment particles’ inherent color, affecting how light is absorbed and scattered in the water. The red coloration in natural light images is generally proportional to the amount of suspended sand or silt in the water.

Additionally, suspended sediment significantly affects how water absorbs and scatters near-infrared light. Clean water absorbs almost all NIR light radiation (> 700 nm), meaning clear water appears dark or black in NIR images [20]. When sediment is present, NIR light is scattered by suspended particles, increasing water-leaving reflectance. Sediment-laden waters thus appear brighter in NIR imagery, particularly in mineral-rich environments. This spectral behavior forms the basis of many remote sensing approaches to estimating SSCs [5,20,23]. Finer particles, such as clay- and organic-rich sediments, scatter less but can still increase NIR reflectance depending on the concentration. Organic materials and contaminants attached to fine sediment can alter NIR absorption. For example, algae- and organic-rich sediments absorb more NIR, making water appear darker, while mineral sediments (e.g., quartz, iron oxides) reflect more NIR, making turbid water appear brighter. Therefore, the brightness of NIR images results from sand- or silt-sized suspended sediment.

3.2. Machine Learning Models

3.2.1. Selection of ML Models Through Preliminary Analysis

In environmental data analysis, a variety of ML models have been successfully applied to capture the nonlinear relationships in high-dimensional datasets [12], often involving water quality, salinity, and sediment applications [24,25]. In our study, we initially evaluated six models—RF, CNN, Multi-Layer Perceptron (MLP), SVM, KNN, and GBM—on approximately 650 images, since prior research indicates their effectiveness in handling structured and unstructured environmental data.

MLP was implemented following the high accuracy reported by [26] in water quality prediction tasks, although its performance in our sediment prediction task was moderate. KNN and SVM were also included based on their documented utility in water quality classification [7,9] but ultimately underperformed on our complex dataset, consistent with findings that small sample sizes or complex, noisy data can inhibit these methods [27,28].

We also implemented the RF model, which is an ensemble method that builds numerous decision trees to capture complex, nonlinear relationships while ranking feature importance in noisy, high-dimensional data [12]. The CNN model was implemented as ResNet50, which leverages deep residual learning to extract complex spatial features from image data [29], and CNNs have also shown effectiveness in marine image classification tasks [11]. However, they typically require large datasets to perform effectively. The MLP is a fully connected neural network with hidden layers designed to model nonlinear relationships in structured data; in our study, we implemented it with three hidden layers to enhance learning capacity, inspired by its successful application in water quality classification [26]. We also used the SVM with an RBF kernel to capture nonlinearities in high-dimensional spaces, a method known for its sensitivity to hyperparameter tuning and computational demands. SVM has also been applied successfully in water quality classification tasks [9]. In addition, we implemented the KNN model, a straightforward distance-based classifier that relies on proximity but often struggles with noise and high dimensionality [7,27], and the GBM, an iterative method that builds decision trees to minimize bias and variance, achieving superior accuracy in complex, multi-dimensional environmental datasets [10,13]. The key parameters used in these models are listed in

Table 1. Key parameters of models tested in the preliminary analysis.

Model	Key Parameters
RF	Estimators: 1200–2000; Max depth: 25–35 or unrestricted; Min samples split: 2–6; Min samples per leaf: 1–3; Max features: all or “sqrt”; Min impurity decrease: 0.0–0.005; Bootstrap: Disabled
CNN	Layers: 3 convolutional (filter sizes: 32, 64, 128) with ReLU; Max-pooling layers; Learning rate: 0.001; Dropout: 0.2–0.5; Optimizer: Adam
MLP	Hidden layers: (128, 64, 32 nodes); Learning rate: 0.001; Activation: ReLU; Regularization term (α): 0.001
SVM	Kernel: Radial Basis Function (RBF); Regularization term (C); Kernel coefficient (γ); Tolerance (ϵ), tuned via RandomizedSearchCV
KNN	Number of neighbors: 3–15; Distance metrics: Euclidean, Manhattan
GBM	Estimators: 1400; Learning rate: 0.05; Maximum tree depth: 20; Subsample ratio: 1.0

.

For natural light images, six features were extracted—mean values of the red, green, and blue channels, red reflectance, time of image capture, and temperature—while NIR images were processed to derive GLCM texture features (contrast, homogeneity, entropy, energy, and dissimilarity) in addition to time of image capture and temperature.

An initial calibration using 100% of the training data allowed us to assess performance using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), R² Score, and Kling–Gupta Efficiency (KGE). These results supported a focus on two methods—RFR and GBR—which best handled noisy data and complex feature interactions [10,12,24].

While models such as CNN, MLP, SVM, and KNN showed promise in theory, their performance was hampered by practical limitations in our dataset. CNNs, for instance, require large datasets to capture spatial patterns reliably, and their computational overhead proved prohibitive given our limited image diversity [11]. Similarly, MLPs, despite high accuracy reported in water quality studies [26], yielded higher error rates in our experiments, likely due to their sensitivity to the limited complexity of the dataset and lack of embedded physical constraints, also likely due to their black-box nature and limited interpretability, which [28] suggests can hinder mechanistic insight in water quality modeling. SVMs, while effective in capturing nonlinear relationships, suffered from computational complexity and inconsistent tuning outcomes, and KNN’s simplistic distance-based approach failed to generalize well in the presence of noisy, high-dimensional features [7,10]. Consequently, our results consistently favored ensemble methods, prompting a focus on RF and GBM due to their robustness, capacity to handle high-dimensional data, and ability to rank feature importance, factors that have been highlighted in environmental modeling studies [12,28].

Our preliminary investigation suggested that RF and GBM outperform other models, a finding supported by [12,13,24], who point out that the selected models handle high-dimensional, noisy data and capture complex feature interactions better than other models. These methods showed comparatively lower error rates in our sediment concentration predictions, motivating the choice of RF and GBM for further analysis. However, the preliminary analysis was conducted on a relatively small dataset, limiting the models’ generalizability. To address this, we collected a significantly more extensive set of images.

The preliminary dataset consisted of approximately 650 images, which included both RGB and NIR images captured under controlled laboratory conditions. These images represented SSCs ranging from 1000 ppm to roughly 110,000 ppm. To improve model robustness and better capture high-concentration variability, the dataset was later expanded to approximately 1300 images. This extended dataset increased the sediment concentration range to approximately 160,000 ppm. The enhancement allowed for more reliable model training, validation, and generalization across the full concentration spectrum.

3.2.2. Model Set-Up—Fine Tuning of Selected Models

Based on our preliminary analysis, the final four models were implemented as follows:

• For Natural Light Images:

Models were trained using RGB mean values, red reflectance, image capture time (extracted from filenames), and temperature. Preprocessing involved extracting average RGB values and red reflectance from each image, incorporating time and temperature as additional features, standardizing the feature set using z-score normalization, and applying log transformation to the target sediment concentration. Hyperparameter tuning was conducted using GridSearchCV with 5- or 9-fold cross-validation, following practices adopted in prior water quality modeling studies [10,13,30].

• For NIR Images:

Models used GLCM texture features derived from NIR images, combined with time and temperature. As described previously, GLCM-based texture features were used to represent spatial variations in NIR images [31]. However, there are a large number of (GLCM) texture features (72 features for RFR and 112 features for GBR). This information is summarized in

Table 2. Summary of model inputs, feature types, and output variable across RGB and NIR image-based regressors.

Model Type	Model	Input Features	No. of Input Features	Output Variable	Modeling Type
RGB	RFR	Mean Red, Mean Green, Mean Blue, Red Reflectance, Time of Capture, Temperature	6	Suspended Sediment Concentration (SSC)	Regression
RGB	GBR	Mean Red, Mean Green, Mean Blue, Red Reflectance, Time of Capture, Temperature	6	Suspended Sediment Concentration (SSC)	Regression
NIR	RFR	GLCM Texture Features (Contrast, Homogeneity, Energy, Correlation, Dissimilarity, Entropy) from 3 distances × 4 angles; Time; Temperature; Data Augmentation	74	Suspended Sediment Concentration (SSC)	Regression
NIR	GBR	GLCM Texture Features (Contrast, Dissimilarity, Homogeneity, Energy, Correlation, ASM, Entropy) from 4 distances × 4 angles; Time; Temperature; Data Augmentation	114	Suspended Sediment Concentration (SSC)	Regression

. The use of GLCM for environmental image analysis, particularly for distinguishing water bodies, has been demonstrated in remote sensing contexts such as SAR-based segmentation [32]. To address limitations in dataset size, basic augmentation techniques were employed, consistent with practices in image-based classification studies using deep learning (e.g., [11]). Model training included GridSearchCV with train–test splits and cross-validation, following best practices in ensemble model optimization [10,13]. Hyperparameter tuning, particularly of estimators, learning rate, and depth, enhanced performance in both GBR and RFR [12,24].

To visualize how each input feature relates to SSC in natural light images, Figure 2 shows scatter plots comparing SSCs with average RGB values, red reflectance, temperature, and time. While no strong linear patterns emerge, higher SSCs generally correspond to increased RGB intensity—consistent with more turbid water appearing more saturated. Red reflectance also shows denser clustering around ~0.35–0.4 at elevated concentrations. In contrast, time and temperature exhibit no discernible trends but were retained to account for potential variability in lighting and environmental conditions. For NIR images, we did not present a similar analysis due to the high dimensionality of GLCM texture features (72 for RFR, 112 for GBR), which are better assessed through model-driven feature importance rather than raw scatter plots.

3.2.3. Feature Selection and Model Optimization

The time of image capture was included as a feature in both the natural light and NIR models to account for temporal variation in lighting conditions throughout the day, as light reflectance is known to change with solar angle and time [20]. These complementary color and temperature features were found to improve model performance and interpretability in our machine learning framework. With expanded image collection and NIR-specific data augmentation, hyperparameter settings for RF and GBM were reassessed to account for increased sample size and feature diversity. These adjustments led to more robust model tuning and performance improvements [10,12,13,14]. The increased data volume and diversity allowed for more robust RF and GBM model tuning. GridSearchCV was conducted again to optimize key parameters such as the number of estimators, depth constraints, and feature selection strategies to accommodate the expanded dataset. For natural light images, the final feature set consists of the average red, green, and blue values, red reflectance, time of capture, and temperature. In contrast, we extracted gray-level co-occurrence matrix (GLCM) texture features for NIR images, which were combined with time of capture and temperature to enhance model performance.

For the GBM, GLCM texture features were extracted using four distances ([1, 2, 5, 10]) and four angles ([0, π/4, π/2, 3π/4]), yielding sixteen distance–angle combinations. From each combination, six standard texture properties—contrast, dissimilarity, homogeneity, energy, correlation, and ASM—were computed, along with entropy, resulting in seven features per pair. This produced a total of 112 GLCM-based features [31]. By appending the time of capture and temperature, the final feature vector included 114 features.

For the RF model with NIR images, GLCM features were computed using three distances ([1, 5, 10]) and four angles ([0, π/4, π/2, π]), resulting in twelve combinations. For each, 6 texture properties and entropy were calculated, totaling 72 GLCM features [31]. Including time and temperature yielded a final feature vector of 74 features. This difference in GLCM configuration allowed the RF model to maintain a more tractable feature space while still capturing representative texture patterns from the imagery.

To enhance model learning from the NIR images, we applied data augmentation techniques—including horizontal and vertical flips, 90° and 180° rotations, slight random rotations, and noise injection—to increase variability and improve generalization. Although the RGB and NIR datasets contained the same number of images, augmentation was applied only to NIR images due to their lower feature diversity and reliance on texture-based inputs. We validated the models using 5-fold or 9-fold cross-validation [10] for natural light images, which provided a more robust assessment than a fixed 80–20 training/validation split. A 70–30 train–test split was used for NIR images, supplemented with 5-fold cross-validation to ensure reliable evaluation. The use of near-infrared spectral data in sediment analysis has been previously validated in image-based environmental studies [5,6].

Hyperparameter tuning for RF and GBM models was conducted using GridSearchCV, systematically optimizing key parameters. The RF model was tuned for the number of estimators, maximum tree depth, minimum samples required for splits and leaves, feature selection strategies (all features vs. “sqrt”), bootstrap settings, and minimum impurity decrease. Similarly, for GBM, tuning focused on the number of estimators, learning rate, maximum tree depth, subsample ratio, and minimum number of samples for splits and leaves.

Table 3. Key model parameters in the final analysis.

Model	Dataset	Features	Key Hyperparameters	Additional Details
RF	RGB	Avg. red, avg. green, avg. blue, red reflectance, time, temperature	n_estimators: [1200, 1500, 1800, 2000]; max_depth: [25, 30, 35, None]; min_samples_split: [2, 4, 6]; min_samples_leaf: [1, 2, 3]; max_features: [None, “sqrt”]; bootstrap: False; min_impurity_decrease: [0.0, 0.002, 0.005]	6-fold CV using GridSearchCV; StandardScaler applied
GBM	RGB	Avg. red, avg. green, avg. blue, red reflectance, time, temperature	n_estimators: [1400, 1600]; learning_rate: [0.03, 0.05]; max_depth: [18, 20]; min_samples_split: [2]; min_samples_leaf: [1]; subsample: 1.0; max_features: [None]	10-fold CV using GridSearchCV; StandardScaler applied; log-transformed and min–max scaled target
RF	NIR	GLCM features (72 features) + time, temperature	n_estimators: 200; max_depth: None; min_samples_split: 5; min_samples_leaf: 1; max_features: “log2”; random_state: 42	Data augmentation applied; train–test split (70–30)
GBM	NIR	GLCM features (112 features) + time, temperature	n_estimators: [100, 300, 500]; learning_rate: [0.01, 0.05, 0.1]; max_depth: [3, 4, 5]; min_samples_split: [2, 5, 10]; min_samples_leaf: [1, 2, 4]	Data augmentation applied; train–test split (70–30); 5-fold CV via GridSearchCV

demonstrates the models and their parameters. These optimizations enhanced model performance metrics by reducing the Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) while increasing the coefficient of determination (R²) and Kling–Gupta Efficiency (KGE) [33] within the scope of our dataset. The KGE combines three components into a single metric:

$$KGE=1-\sqrt{{\left(r-1\right)}^2+{\left(\beta -1\right)}^2+{\left(\gamma -1\right)}^2}$$

(2)

In which r is the Pearson correlation coefficient between observed and simulated values; $\beta ={\displaystyle \frac{{\mu }_s}{{\mu }_o}}$ is the bias ratio, in which μ_s is the mean of simulated values and μ_o is the mean of observed values; and $\gamma ={\displaystyle \frac{{CV}_s}{{CV}_o}}={\displaystyle \frac{{\sigma }_s/{\mu }_s}{{\sigma }_o/{\mu }_o}}$, i.e., the variability ratio defined as the coefficient of variability in simulated vs. observed, where σ is the standard deviation. KGE = 1 means perfect agreement between observed and simulated values, KGE < 1 means deviation from perfect agreement, and KGE < 0 means the model performs worse than the mean of observations.

3.2.4. Computational Environment

All machine learning experiments were conducted on the University of Arizona’s HPC cluster using the GPU standard queue. We utilized single nodes with 16 CPU tasks (6 GB memory per CPU), two Pascal GPUs, and a 4 h wall time limit. This configuration is consistent with HPC-based training strategies adopted in large-scale hydrologic and water quality modeling [34,35], providing sufficient resources for GridSearchCV and data augmentation. Training and validation typically took between 25 to 45 min under these settings.

4. Results

This study evaluated the performance of six machine learning models for sediment concentration prediction using natural light and near-infrared light images. Model performance was assessed using comprehensive metrics, including RMSE, MAE, R², and KGE. These metrics were selected for their ability to capture statistical accuracy and physical consistency, particularly focusing on KGE and R² as critical indicators in hydrologic prediction tasks [34,35].

4.1. Preliminary Model Evaluation

In the preliminary analysis, six machine learning models were evaluated using features from natural light images (mean R, G, B values and red reflectance) to assess their predictive performance for SSC. The models included RFR, GBR, MLP, CNN, SVM, and KNN. The modeling results were evaluated using the percentage of results that have a normalized error of less than 30%. The normalized error is defined as the ratio of the absolute error (i.e., simulation–observation) vs. the observation.

As shown in

Table 4. Summary of preliminary results.

Model	Image Type	% Within 30% Relative Error
RFR	Natural Light (RGB)	83.56%
GBR	Natural Light (RGB)	85.84%
MLP	Natural Light (RGB)	53.88%
CNN	Natural Light (RGB)	32%
SVR	Natural Light (RGB)	49.77%
KNN	Natural Light (RGB)	47.95%
RFR	Infrared (GLCM)	74.83%
GBR	Infrared (GLCM)	88.81%

, both RFR and GBR outperformed the other models, consistent with prior studies recommending tree-based ensemble methods for structured environmental data [12]. Although CNN and MLP achieved good performance in prior studies [11,26], their effectiveness can be limited in practice by small training datasets and sensitivity to hyperparameters. In previous studies, KNN and SVM models underperformed relative to ensemble and Artificial Neural Network-based models in water quality prediction tasks [7].

Based on these results, RFR and GBR were selected for further evaluation on NIR-based feature sets to validate their generalizability.

4.2. Final Model Evaluation

After selecting RFR and GBR as the best-performing models [12,13], we performed hyperparameter tuning on an extended dataset of both natural RGB and NIR images, following the tuning and validation strategies demonstrated in [10,13,30]. Either cross-validation or train–test splits were applied. This step addressed the distinction between color-based features and texture descriptors by utilizing a GLCM-based method originally introduced in [31], while also incorporating temperature data [27] and image capture time to enhance the accuracy of SSC predictions. To evaluate the robustness and generalizability of each model, we computed the error bars for key performance metrics—such as RMSE and R²—by reporting the mean and standard deviation across folds or repeated test splits. These error bars capture the variability in model predictions across different data subsets and provide a more comprehensive understanding of model stability under varying conditions.

Table 5. Cross-validated and repeated train–test split model performance (mean ± standard deviation) for RMSE and R² across four model–image combinations.

Model	Image Type	RMSE (ppm)	R2
RFR	RGB	22,091.14 ± 4541.42	0.5347 ± 0.1855
GBR	RGB	20,695.49 ± 5633.33	0.5878 ± 0.1802
RFR	NIR	13,702.18 ± 977.66	0.8223 ± 0.0226
GBR	NIR	10,865.62 ± 633.55	0.8884 ± 0.0146

summarizes the results across repeated test splits as well as folds, which reflects performance variability across multiple data splits.

The results in Table 5 and

Table 6. Summary of final results.

Model	Image Type	RMSE	MAE	R2	KGE	% Within 10% Error	% Within 20% Error	% Within 30% Error
RFR	RGB	22,562.77	10,063.95	0.55	0.76	52.92%	63.62%	72.18%
GBR	RGB	21,466.45	10,100.58	0.59	0.77	49.42%	62.06%	72.96%
RFR	Infrared (GLCM)	13,471.50	6781.57	0.82	0.78	53.59%	70.90%	78.72%
GBR	Infrared (GLCM)	9885.79	5151.68	0.90	0.89	60.51%	74.87%	83.08%

show that GBR consistently outperformed RFR, particularly on infrared images. While [13] focused on water quality prediction using environmental variables, their findings similarly demonstrate the superior performance of boosting-based methods over random forests when handling complex, multivariate data. The yellow and brownish color in natural rivers and streams is due to high concentrations. As the concentration increases, the color tends to become more saturated and shifts toward deeper hues, indicating lower reflectance in visible bands. Specifically, GBR using NIR images achieved nearly 83% of predictions within a 30% relative error, slightly higher than RFR with NIR images at 79%.

Scatter plots were generated to compare observed and predicted sediment concentrations for each of the four final models: RFR and GBR using natural light images (shown in Figure 3 and Figure 4, respectively), and RFR and GBR using NIR images (shown in Figure 5 and Figure 6, respectively). It’s obvious that the majority of predicted SSC matched the observed.

Additionally, Figure 7 illustrates the relative errors between predicted and measured SSCs using the RFR model with both natural light and NIR images. With natural light images, 80.73% of the data fell within a 50% relative error, 72.18% within 30%, and 52.9% within 10%. When using NIR images, the prediction accuracy improved, with 86.3% of the results within 50% error, 78.72% within 30%, and 53.6% within 10%. These findings indicate that using NIR images enhances the predictive performance of the RFR model.

Figure 8 presents the relative errors of the GBR model using both natural light and NIR images. When using natural light images, 80.74% of the predictions fell within a 50% error margin, 73% within 30%, and 49.4% within 10%. In comparison, the predictions using NIR images show improved accuracy, with 88.6% within 50% error, 83.1% within 30%, and 60.5% within 10%. These results confirm that NIR images yield significantly better performance, and overall, the GBR model outperforms the RFR models.

4.3. Discussion

GBR consistently delivered higher predictive skill than RFR for both visible light and near-infrared imagery. Under natural light, GBR achieved 72.96% of predictions within a 30% relative error band, slightly improving on RFR’s 72.18%. This small but consistent gain reflects the kind of improvement reported by [12] for boosting over bagging ensembles in hydrological applications. When NIR imagery was used, GBR’s advantage became more pronounced. GBR achieved 83.08% of predictions within the 30% error band, compared with 78.72% for RFR. This larger margin suggests that GBR’s stage-wise additive learning strategy is well suited for modeling the nonlinear, high-dimensional relationships captured by GLCM texture features [13,32]. Hydrologically focused metrics, including R² and KGE, further corroborated GBR’s superior performance and support its suitability for image-based sediment monitoring [34,36]. Although based on a modest laboratory dataset, these results highlight the promise of combining spectral imagery with boosting algorithms for scalable, non-intrusive SSC assessment.

5. Conclusions

This study demonstrates that machine learning models, especially GBR and RFR, can provide accurate estimates of suspended sediment concentrations from RGB and NIR imagery when combined with environmental and temporal covariates. GBR’s stronger performance, especially under NIR illumination, where it reached 83.08% within the 30% error threshold, underscores its value for non-invasive sediment monitoring. This result highlights GBR’s strength in modeling the complex texture–concentration relationships captured in near-infrared images. These findings suggest that combining spectral data with boosting methods offers a robust, scalable, and non-invasive pathway for real-time sediment monitoring. However, as the experiments were conducted in a controlled flume with a moderate sample size, larger field datasets and in situ validation are needed to confirm transferability. Future work should expand the image library, incorporate additional environmental drivers such as turbidity and irradiance, and assess model performance in natural river settings to enhance robustness and generalizability.