In Situ Hyperspectral Reflectance Sensing for Mixed Water Quality Monitoring: Insights from the RUT Agricultural Irrigation District

Abstract

Water-quality monitoring in agricultural irrigation systems is challenging due to the dynamic and heterogeneous nature of mixed water sources, which complicates traditional and remote sensing-based assessment methods. Traditional water quality monitoring relies on water sampling and laboratory analysis, which can be time-consuming, labor-intensive, and spatially limited. In situ hyperspectral reflectance sensing (HRS) presents a promising alternative, offering high-resolution, non-invasive monitoring capabilities. However, applying HRS in mixed-water environments—where served-water effluent, precipitation, and natural river water converge—presents significant challenges due to variability in water composition and environmental conditions. While HRS has been widely explored in controlled or homogeneous water bodies, its application in highly dynamic agricultural mixed-water systems remains understudied. This study addresses this gap by evaluating the relationships between in situ hyperspectral data (450–900 nm) and key water-quality parameters—pH, turbidity, nitrates, and chlorophyll-a—across three campaigns in a Colombian tropical agricultural irrigation system. A Pearson’s correlation analysis revealed the strongest spectral associations for nitrates, with positive correlations at 500 nm (r ≈ 0.76) and 700 nm (r ≈ 0.85) and negative correlations in the near-infrared (850 nm, r ≈ −0.88). Conversely, the pH exhibited weak and diffuse correlations, with a maximum of r ≈ 0.51. Despite their optical activity, turbidity and chlorophyll-a showed unexpectedly weak correlations, likely due to the optical complexity of the mixed water matrix. Random Forest regression identified key spectral regions for each parameter, yet model performance was limited, with R² values ranging from 0.51 (pH) to −1.30 (chlorophyll-a), and RMSE values between 0.41 and 1.51, reflecting the challenges of predictive modeling in spatially and temporally heterogeneous wastewater systems. Despite these challenges, this study establishes a baseline for future hyperspectral applications in complex agricultural water monitoring and highlights critical spectral regions for further investigation. To improve the feasibility of HRS in mixed-water assessments, future research should focus on enhancing data-preprocessing techniques, integrating complementary sensing modalities, and refining predictive models to better account for environmental variability.

1. Introduction

Water-quality monitoring is essential for safeguarding ecosystems, public health, and sustainable agricultural practices. Effective monitoring helps prevent pollution, ensures resource viability, and supports key sectors, such as drinking-water supply, irrigation, and industry [1] (Zheng et al., 2024). However, agricultural irrigation systems, especially in developing regions, face significant challenges due to urban expansion, inadequate wastewater treatment infrastructure, and reliance on untreated or minimally treated water. These issues introduce contaminants, including heavy metals, pathogens, and excessive nutrient loads, which can negatively impact soil health, crop productivity, and food safety [2,3,4] (Gholizadeh et al., 2016; Jaramillo & Restrepo, 2017; Rahat et al., 2023).

Wastewater reuse in agriculture, while addressing freshwater scarcity, presents risks associated with pollutants that may compromise food security and environmental sustainability [5] (Helmecke et al., 2020). Traditional water-quality monitoring methods rely on laboratory analysis of collected samples, providing accurate but time-consuming and spatially limited results. These conventional approaches struggle to offer real-time, large-scale assessments needed for effective water resource management [6] (Yang et al., 2022).

Hyperspectral reflectance sensing (HRS) has emerged as a promising alternative, enabling high-resolution, non-invasive water-quality monitoring. Unlike conventional multispectral imaging, which captures reflectance in discrete bands, HRS provides a continuous spectral profile, enhancing the detection of subtle changes in water-quality indicators [7] (de Oliveira et al., 2020). This technology facilitates the identification of optically active parameters, such as turbidity, chlorophyll-a, total suspended solids (TSS), and colored dissolved organic matter (CDOM), across the visible (VIS), near-infrared (NIR), and shortwave infrared (SWIR) regions [8,9] (Rahul et al., 2023; Wei et al., 2020). However, applying HRS to wastewater-impacted agricultural irrigation systems remains complex due to the highly variable composition of mixed-water sources, including served-water effluents, precipitation, and river water.

This study investigates the potential of in situ HRS (450–900 nm) for monitoring water-quality parameters—pH, turbidity, nitrates, and chlorophyll-a—across three field campaigns in a tropical agricultural irrigation system. By analyzing spectral responses, we aim to determine the feasibility of HRS in capturing water-quality dynamics in complex mixed-water systems and assess the challenges associated with its application in such environments. systems.

2. Related Work

The use of remote sensing for water-quality assessment has expanded significantly, particularly with the application of hyperspectral techniques. Research has shown that HRS can effectively detect optically active parameters, including chlorophyll-a [10] (Uudeberg et al., 2020), phycocyanin [11] (Li et al., 2010), turbidity [12,13] (Cui et al., 2022; Yim et al., 2020), and CDOM [14] (Arabi et al., 2020). Additionally, some studies have explored the feasibility of detecting optically inactive parameters, such as ammonium nitrogen (NH₄ -N) [15] (Cao et al., 2022), total phosphorus (TP) [16] (Sun et al., 2022), nitrate nitrogen [17] (Al-Shaibah et al., 2021), and chemical oxygen demand (COD) [16] (Sun et al., 2022), by leveraging correlations with optically active substances.

Machine learning (ML) techniques have become increasingly valuable in water-quality modeling, particularly when integrated with multi- and hyperspectral remote sensing to enhance prediction accuracy. A study in Nansi Lake, China, combined Sentinel-2 imagery with a Stacking Model that integrated eight ML algorithms, improving prediction accuracy by 12% over XGBoost and mitigating high-value underestimation and low-value overestimation. Key spectral indices for chlorophyll-a and turbidity estimation, including Rrs(705)/Rrs(665) and the Normalized Difference Turbidity Index, were identified [18] (J. Zhang et al., 2024). Similarly, a study conducted in marine areas near agricultural lands employed a low-cost RGB-based optical sensor with ML to quantify and classify turbidity. Using 64 light combinations, an Exponential Gaussian Process Regression model achieved an R² of 0.979 for turbidity quantification, while a Fine K-Nearest-Neighbor classification attained 91.23% accuracy, demonstrating the feasibility of low-cost optical sensors for water-quality assessment [19] (Parra et al., 2024).

Remote sensing techniques have been widely applied to inland and coastal waters. Studies have demonstrated strong predictive relationships between spectral data and key water-quality parameters in lakes, reservoirs, and rivers. For example, research in China’s Lake Taihu, Liangxi River, and Fuchunjiang Reservoir validated the use of proximal hyperspectral sensing to estimate TN, TP, and COD using spectral signatures in the red and NIR ranges (700–800 nm), achieving predictive accuracies exceeding 90% with machine learning models [16] (Sun et al., 2022). Similarly, HRS in Singapore reservoirs successfully monitored chlorophyll-a and turbidity, with R² values above 0.8 [20] (Liew et al., 2019).

The integration of HRS with satellite-based remote sensing has also proven effective in large-scale water quality monitoring. Studies in the Dutch Wadden Sea combined in situ hyperspectral data with MERIS, MSI, and OLCI satellite observations to track long-term variations in chlorophyll-a, suspended particulate matter (SPM), and CDOM [14] (Arabi et al., 2020). ML approaches, such as Mixture Density Networks, have been used to harmonize Sentinel-2 and Sentinel-3 data for chlorophyll-a retrievals [21] (Pahlevan et al., 2020).

Unmanned Aerial Vehicles (UAVs) equipped with hyperspectral sensors have further enhanced water-quality monitoring. Studies in China’s Miyun Reservoir and Beigong Reservoir achieved highly accurate TN and chlorophyll-a retrievals (R² = 0.99 and R² = 0.96, respectively) using machine learning models [22,23] (Lu et al., 2021; Qun’ou et al., 2021). However, UAV-based HRS remains costly and subject to operational constraints, including flight regulations and atmospheric noise.

While existing research has demonstrated the feasibility of HRS for various water-quality applications, few studies have explored its performance in mixed-water agricultural irrigation systems—highly dynamic environments where water sources, flow conditions, and contaminant loads fluctuate significantly. Addressing this knowledge gap, our study applies in situ HRS to assess its potential for monitoring key water-quality indicators in a tropical agricultural irrigation canal, providing a baseline for future research and methodological improvements.

3. Methods

3.1. Study Area

La Unión municipality is in the northern region of the Valle del Cauca department in Colombia. It forms part of the Roldanillo-Unión-Toro (RUT) agricultural irrigation district, which is recognized as one of the most significant irrigation areas in the country (Figure 1). This municipality serves as a vital agricultural hub, contributing over 50% of the food crop production within the Valle del Cauca Department (southwest Colombia). The RUT district encompasses approximately 10,256 hectares, managed by the RUT Users Association (ASORUT) (La Unión, Colombia), which oversees water access for the irrigation of various crops.

The region’s agricultural landscape is dominated by sugarcane, maize, guava, and grapes, along with a variety of other fruits and vegetables. To determine an optimal sampling site, initial exploratory campaigns were conducted at wastewater outlet pipes in Roldanillo (south) and Toro (north). Based on these assessments, La Unión’s wastewater outlet pipe was selected as the primary sampling location. At this site, wastewater from the main outlet pipe mixes with water from the Cauca River, which is contained within an open canal. This canal serves a dual purpose: functioning as an irrigation canal during the dry season and as a drainage canal in the rainy season. Sampling was conducted 20 m north of the outlet pipe (Figure 1), where conditions were favorable due to accessibility, adequate canal width, and stable water flow.

The selected site exhibited minimal interference from floating debris or branches and avoided varying light conditions that could lead to reflectance anomalies (Figure 2). Following site selection, three field campaigns were conducted to gather data on contrasting reflectance and water-quality parameters. Campaigns occurred from October to December 2024 to capture the variability in climate and weather in the sampling-point area. It was not possible to run a monthly collection of data due to a limited financial capacity. The data collection involved in situ spectral measurements using two spectrometers and simultaneous water sampling at the designated point. This comprehensive approach aimed to enhance the accuracy and reliability of our findings regarding mixed-water-quality monitoring in agricultural irrigation contexts.

3.2. Water-Quality Parameters Selection

Four water-quality parameters were selected for this study: nitrates, pH, turbidity, and chlorophyll-a. The selection criteria were based on two fundamental principles: first, they must accurately represent and reflect the characteristics of water pollution; second, they should be widely recognized and utilized in routine water monitoring, supported by robust scientific evidence regarding their effectiveness and practicality. The selected parameters were classified into optically active (turbidity and chlorophyll-a) and optically inactive groups (nitrates and pH). This classification is crucial for understanding the spectral responses associated with different water-quality characteristics (

Table 1. List of selected water parameters used in this research based on their optical activity.

Optically active	Chlorophyll-a	[15] (Cao et al., 2022)
	Turbidity	[13] (Yim et al., 2020)
Optically inactive	pH	[14] (Arabi et al., 2020)
	Nitrate nitrogen (NO3-N)	[17] (Al-Shaibah et al., 2021)

).

The data collection was conducted over three campaigns that encompassed a diverse range of environmental conditions, including variations in light levels, flow regimes, and seasonal changes (

Table 2. Environmental conditions during each campaign (MAT: mean air temperature; WFI: water flow intensity).

Campaign	Date	Cloudiness (%)	MAT (°C)	WFI
1	15 October 2024	~20%	29–30	Medium
2	20 November 2024	~30%	28–29	Low
3	9 December 2024	~20%	30–31	Medium

, Figure 1). As a general reference, Figure 1 shows a “low” water flow intensity. This approach facilitated the acquisition of a more robust dataset, enhancing the reliability of our findings. By integrating both optically active and inactive parameters, this research aimed to capture a comprehensive picture of water-quality dynamics in the study area.

3.3. Data Used

3.3.1. Water Sampling and Laboratory Analysis

We conducted simultaneous in situ measurements of HRS data and water sampling to evaluate key water-quality parameters. A total of 40 water samples were collected during each campaign, with samples obtained from depths ranging from 30 to 50 cm below the open canal’s surface. These samples were promptly stored, refrigerated and transported to the laboratory for analysis of pH, turbidity, nitrates, and chlorophyll-a.

The open canal’s depth varied significantly, ranging from 1 m during the dry season to over 2 m at its deepest point during the rainy season. Given the nature of the mixed waters, the sampling depths exceeded the transparency of the water column, categorizing them as non-optically shallow (Figure 2).

The determination of nitrates was performed using the DIN 38405 D9-2 method, while the water pH was measured using the electrometric method (standard method 4500-H + B), and turbidity was assessed using the nephelometric method (standard method 2130B). The determination of chlorophyll-a was conducted following Method 10200 H from the 24th edition of the Standard Methods for the Examination of Water and Wastewater [24] (Lipps et al., 2023). The process involved filtering water samples to capture phytoplankton, extracting chlorophyll using an aqueous acetone solution, and macerating the filter. The extract was clarified through centrifugation and subsequently analyzed spectrophotometrically at specific wavelengths (664, 647, and 630 nm). Chlorophyll-a concentration was calculated based on measured absorbances and formulas derived from extinction coefficients.

3.3.2. Reflectance Measurement and Pre-Treatment

At the sampling site, in situ radiance data were measured 50 cm above the water surface immediately following water sampling using two spectrometers. This proximity of the sensors to the water surface ensures high-resolution spectral data, effectively minimizing atmospheric interferences, enhancing signal-to-noise ratios, and enabling precise spectral reflectance measurements [16] (Sun et al., 2022).

The two STS spectrometers used in this study consisted of an STS:VIS (336–822 nm) and an STS:NIR (632–1123 nm) unit (Ocean View, Dunedin, FL, USA), both with an optical resolution of 1.5 nm ± 0.05 nm and 1024 pixels. Each spectrometer was coupled to a PC and equipped with a Gershun tube (aperture of 8°) for optical control.

Measurements were conducted within one hour of solar noon to optimize light conditions. Prior to spectral data collection, a field calibration was performed using a spectralon diffuse reference panel with 99% reflectivity across the 400–1500 nm spectral range (Ocean View, Dunedin, FL, USA). The spectralon was used to monitor reference irradiance, minimize saturation, and account for changes in lighting conditions, ensuring measurement consistency [25] (Reyes-Trujillo et al., 2021). The target was oriented at an angle of 40° to 45° from nadir to effectively capture sunlight; each spectrometer was calibrated according to the specific brightness conditions at that time. The dark spectrum was achieved by completely obstructing each spectrometer with a black cap. These bright and dark conditions were utilized by the software to calibrate for natural light variations. The acquisition of spectral data was completed within 3 to 4 min per campaign, ensuring minimal variation in environmental conditions. This short acquisition period helped mitigate spectral noise caused by fluctuations in natural solar luminosity, enhancing data consistency and reliability. A total of 45 spectral signatures were recorded per campaign, comprising 45 NIR and 45 VIS measurements. This approach addressed factors such as fluctuations in ambient temperature and lighting conditions, variations in light source intensity, baseline drift in the spectrometer, and physical alterations to the experimental setup.

Pre-processing of the raw data was conducted using Ocean View spectroscopy software version 2.015 (Ocean View, Dunedin, FL, United States). The combined HRS raw data (336–1123 nm) were subsequently narrowed down to a range of 450–900 nm, following methodologies established in previous studies [9,14,16] (Arabi et al., 2020; Sun et al., 2022; Wei et al., 2020). Spectral regions below 450 nm and above 900 nm were excluded due to high levels of noise and artifacts, such as anomalous peaks caused by reflectance anomalies. Pre-processing techniques, such as smoothing, de-trending, and removal of noise or artifacts from reflectance spectra, are critical to improve the quality of the data [10] (Uudeberg et al., 2020).

The selected wavelength range encompasses regions sensitive to the chosen water parameters [15,16,17] (Al-Shaibah et al., 2021; Cao et al., 2022; Sun et al., 2022), thereby providing a comprehensive spectral dataset for subsequent statistical analyses.

3.4. Statistics

3.4.1. Preprocessing and Pearson Correlations

Outlier detection was conducted using reflectance values outside the physical range of 0% to 40%. Subsequently, a boxplot-based analysis was applied to identify and remove additional extreme values.

The cleaned datasets underwent a detailed Pearson’s correlation analysis to evaluate the relationships between spectral reflectance and water-quality parameters—nitrates, pH, turbidity, and chlorophyll-a. Following established protocols in spectroscopy research, this analysis provided foundational insights into potential linear correlations across wavelengths [16,25] (Reyes-Trujillo et al., 2021; Sun et al., 2022).

The Pearson’s correlation analysis was structured into three stages. In the first stage, a Python version 3.10-based visualization tool plotted the correlation coefficients as a function of wavelength, highlighting significant thresholds (0.5 and −0.5) to demarcate regions of notable positive and negative correlations. The correlation values ranged between −0.88 and 0.85, with the strongest relationships observed in the 500–700 nm and 800–900 nm spectral regions. In the second stage, the tool identified the most significant correlations for each parameter, highlighting these on the plots with red markers and annotated values for enhanced interpretability. Finally, a Savitzky–Golay filter (window length = 11, polynomial order = 2) was employed to smooth the trend lines, reducing noise while preserving critical spectral features. This approach facilitated the identification of active spectral regions, where clusters of wavelengths exhibited correlations with specific water-quality parameters. The detailed graphical and statistical evaluation provided by this approach supports the identification of wavelength regions sensitive to specific parameters without relying on complex algorithms to detect initial trends in the data. This forms the basis for subsequent analyses and ensures the effective use of spectroscopic data in water-quality monitoring.

3.4.2. Non-Linear Analysis

A Random Forest (RF) regression model was developed to evaluate the relationship between spectral data and water-quality parameters (nitrates, pH, turbidity, and chlorophyll-a). The predictor dataset (X) was standardized using a z-score transformation (mean = 0, standard deviation = 1) to ensure uniform scaling, and the response dataset (Y) was similarly standardized. The dataset was then split into a training set (80%) and an independent test set (20%) using stratified random sampling to maintain variability across water quality conditions. The split was conducted randomly to ensure a representative distribution of water-quality parameters across both subsets. Cross-validation (k = 5 folds) was applied exclusively to the training set to assess model stability and avoid overfitting. Each RF model was trained with 100 trees (n-estimators = 100), and feature importance was computed based on the mean decrease in impurity (MDI). Model performance was evaluated using root-mean-squared error (RMSE), coefficient of determination (R²), and ratio of performance to deviation (RPD), which quantifies predictive strength. The final model was assessed on the unseen test set, ensuring a robust estimation of generalization performance. This non-linear analysis offered a robust framework for identifying key spectral features associated with water-quality parameters, leveraging RF’s capability to handle high-dimensional datasets and uncover complex relationships that may be overlooked by simpler models.

4. Results

4.1. Water-Quality Parameters Across Campaigns

4.1.1. Campaign 1

The nitrate concentrations in Campaign 1 exhibited relatively low variability, with a mean of 1.51 mg/L (SE = 0.16, SD = 0.7) and a median of 1.3 mg/L, suggesting a slightly left-skewed distribution. The nitrate boxplot (Figure 3a) confirms this stability, with a narrow interquartile range (IQR) and few outliers. The pH values were highly consistent, averaging 7.15 (SE = 0.03, SD = 0.19), reflecting near-neutral conditions. The median (7.12) closely aligned with the mean, supporting the symmetry of the distribution. The pH boxplot (Figure 3b) further demonstrates this stability, with minimal IQRs and no significant outliers, indicating strong buffering capacity in the canal system. Turbidity showed moderate variability, with a mean of 28.9 NTU (SE = 0.41, SD = 2.59). The median (29 NTU) closely matches the mean, indicating a roughly symmetrical distribution. The boxplot (Figure 3c) illustrates relatively low turbidity levels compared to other campaigns, with limited variability and a small number of outliers. Chlorophyll-a exhibited the highest variability among the parameters, with a mean of 249.03 mg/m³ (SE = 19.46, SD = 123.08) and a median of 267.55 mg/m³, suggesting a left-skewed distribution. The boxplot (Figure 3d) highlights this variability, showing a broad IQR and the presence of some extreme values, indicating occasional reductions in chlorophyll-a levels within the canal system.

4.1.2. Campaign 2

The nitrate concentrations during Campaign 2 showed increased variability, with a mean of 1.56 mg/L (SE = 0.27, SD = 0.82) and a median of 1.33 mg/L, indicating a slight positive skew. The boxplot (Figure 3a) reveals a wider IQR compared to other campaigns and notable outliers exceeding 2.5 mg/L, suggesting episodic nitrate inputs, potentially from agricultural runoff or wastewater surges. The pH remained remarkably stable, with a mean of 7.24 (SE = 0.01, SD = 0.06) and an identical median of 7.24, confirming a highly symmetrical and tightly distributed dataset. The boxplot (Figure 3b) supports this finding, showing minimal variability and no significant outliers, reinforcing the system’s buffering capacity. Turbidity exhibited the highest recorded levels across all campaigns, with a mean of 71.63 NTU (SE = 0.83, SD = 5.24) and a median of 70 NTU, indicating significant particulate presence. The boxplot (Figure 3c) illustrates this, with the largest IQR among campaigns and outliers exceeding 80 NTU, suggesting disturbances such as sediment resuspension, localized runoff, or anthropogenic activity. Chlorophyll-a concentrations displayed substantial fluctuations, with a mean of 97.62 mg/m³ (SE = 13.23, SD = 83.69) and a median of 70.4 mg/m³, indicating a right-skewed distribution. The boxplot (Figure 3d) confirms this, showing a narrow IQR and a few extreme values, suggesting a subset of samples with elevated chlorophyll-a concentrations, likely linked to specific environmental or nutrient conditions.

4.1.3. Campaign 3

The nitrate levels in Campaign 3 showed the lowest variability among the campaigns, with a mean of 1.28 mg/L (SE = 0.1, SD = 0.42) and a median of 1.32 mg/L, suggesting a slight left-skewed distribution. The boxplot (Figure 3a) supports this observation, with the narrowest IQR, indicating a more consistent nitrate distribution during this period. The pH remained stable, averaging 7.21 (SE = 0.02, SD = 0.11), with a median of 7.20, confirming a near-symmetrical distribution. The boxplot (Figure 3b) shows minimal IQRs and no outliers, further reinforcing the stability of pH values throughout the study period.

Turbidity levels were lower compared to Campaign 2 but still exhibited moderate variability, with a mean of 42.38 NTU (SE = 0.79, SD = 4.98) and a median of 40 NTU. The boxplot (Figure 3c) highlights slightly increased turbidity compared to Campaign 1, with a moderate IQR and some outliers. The slightly positive skew in the data distribution suggests occasional increases in suspended solids, potentially due to seasonal changes in flow conditions. Chlorophyll-a concentrations were the highest among all campaigns, with a mean of 324.05 mg/m³ (SE = 14.62, SD = 92.49) and a median of 304.32 mg/m³, reflecting a right-skewed distribution. The boxplot (Figure 3d) reveals significant variability, with an extensive IQR and outliers exceeding 400 mg/m³, indicating periodic algal blooms, possibly influenced by nutrient availability and hydrological conditions.

4.1.4. Boxplot of the Hyperspectral Data

Reflectance values in Campaign 1 exhibit a median reflectance of approximately 6.61%, with the widest IQR among the three campaigns (4.35% to 12.47%) (Figure 4). This indicates significant variability in reflectance during this campaign. The whiskers extend asymmetrically, with a greater spread above the median, suggesting a skew toward higher reflectance values. This variability could reflect heterogeneous surface conditions, such as differences in particulate matter, algae, or water surface properties. The asymmetrical whiskers may be influenced by episodic factors or localized disturbances affecting reflectance measurements.

Campaign 2 demonstrates the highest median reflectance, approximately 13.66%, with a relatively narrow IQR (11.83% to 15.22%) (Figure 4). The symmetrical whiskers indicate a balanced distribution of reflectance values, reflecting uniform surface conditions during the sampling period. This consistency suggests a more stable environment, potentially influenced by steady water flow or uniform light conditions. The elevated reflectance median compared to the other campaigns may be attributed to increased surface brightness or particulate matter during this period. The plot also indicates that this campaign presented the highest number of outliers.

Campaign 3 shows the lowest median reflectance, approximately 2.79%, with the narrowest IQR (2.25% to 3.68%) (Figure 4). The symmetrical whiskers reflect minimal variability in reflectance, indicating highly consistent surface conditions during the campaign. The tight clustering of reflectance values suggests uniform environmental factors, such as stable light conditions and consistent water surface properties. The low reflectance values may be indicative of reduced particulate matter or changes in water surface composition, such as decreased turbidity or algal activity.

4.2. Average Reflectance Curves

4.2.1. Campaign 1

Reflectance in the visible spectrum (Figure 5) shows a pronounced peak between 500 and 600 nm, corresponding to green wavelengths. This peak likely results from high reflectance due to the presence of surface materials, such as vegetation or algal content, which are known to reflect strongly in the green region. The variability (observed as oscillations in the blue line) may reflect heterogeneity in surface conditions, including the distribution of particles or biological matter at the water surface. The gradual decline in reflectance beyond 600 nm suggests reduced scattering or absorption effects dominating in the red part of the spectrum, potentially due to increased water absorption or suspended matter that attenuates red light.

Reflectance values in the NIR region are significantly lower and stabilize around 5%. The steep decline after 700 nm is typical of water bodies, where absorption by water molecules in the NIR range dominates over scattering (Figure 5). The stabilization beyond 800 nm indicates a minimal contribution from other scattering sources, such as particulate matter or algae, which are less reflective in this range.

The reduced variability (blue line oscillations) in the NIR region, compared to the visible spectrum, suggests more uniform optical properties across the sampled scans, potentially reflecting homogeneity in water surface composition during the measurements (Figure 5).

The observed reflectance trends align with the expected spectral behavior of water bodies (Figure 5). The peak in the visible green region (500–600 nm) and the decline into the NIR (700–900 nm) suggest that the water surface is influenced by particulate matter, chlorophyll content, and water absorption properties. The reduced variability in the NIR region indicates that water’s absorption properties dominate over scattering effects, masking heterogeneity observed in the visible spectrum.

4.2.2. Campaign 2

In the visible spectrum, the reflectance values display a pronounced peak around 550 nm, corresponding to the green region of the spectrum (Figure 6). This peak is consistent with the typical spectral behavior of surfaces influenced by vegetation or chlorophyll, as green wavelengths are highly reflective due to their minimal absorption by pigments. The gradual decline in reflectance beyond 550 nm into the red region (600–700 nm) is likely caused by increasing absorption from water and suspended particulate matter, including organic and inorganic substances.

The oscillations observed in the blue line (average reflectance) reflect variability across scans, potentially attributable to differences in surface conditions such as the presence of algal biomass, suspended sediments, or variations in illumination during measurements. The red line (smoothed curve) captures the dominant trend, indicating a stable and characteristic pattern for Campaign 2.

In the NIR region, the reflectance stabilizes between 10% and 15%, with less pronounced fluctuations compared to the visible spectrum (Figure 6).

4.2.3. Campaign 3

The visible spectrum shows a pronounced peak in reflectance around 550 nm, corresponding to the green wavelength region (Figure 7). This peak likely reflects the dominance of chlorophyll or other biological materials, which strongly reflect green light. The reflectance gradually decreases beyond 550 nm as absorption by water and suspended particulate matter increases in the red wavelength region (600–700 nm).

Variability in the average reflectance curve (blue line) is evident throughout the visible range, with fluctuations potentially attributed to heterogeneity in water surface conditions, including differences in particulate matter, algal distribution, or variable light conditions during the scans. The smoothed curve (red line) effectively captures the general trend, providing a robust representation of the spectral reflectance pattern.

Reflectance in the NIR region declines sharply after 700 nm and stabilizes around 2–3% reflectance (Figure 7). This trend is consistent with the optical behavior of water, where absorption dominates in the NIR region. The stabilization of reflectance indicates minimal influence from scattering due to particulate matter or algae, which are less reflective in this spectral range. The reduced variability in the NIR region compared to the visible spectrum suggests more uniform optical properties in the sampled scans. This stabilization reflects the intrinsic water absorption characteristics and may indicate minimal disturbances affecting the NIR wavelengths during this campaign.

4.3. Active Spectral Regions

4.3.1. Campaign 1

The correlation plot for nitrates demonstrates varying relationships across the wavelength range, with notable positive peaks around 500 nm (correlation ~0.61) and 700 nm (~0.60) (Figure 8a). These wavelengths correspond to the green and red spectral regions, respectively. The positive correlation in these regions suggests that increased reflectance is associated with higher nitrate concentrations, possibly due to interactions with particulate or dissolved organic matter (DOM) that may enhance reflectance in these bands. Conversely, negative correlations are observed in the NIR region (800–900 nm), with a minimum around 800 nm (−0.74). This negative trend aligns with the dominance of water absorption properties, where nitrate-rich conditions might reduce reflectance due to associated increases in water turbidity or absorption.

For pH, the strongest positive correlations occur in the NIR region (~800–900 nm) (Figure 8b), with peaks exceeding 0.6 near 830 nm. This trend indicates a robust relationship between pH and reflectance in the NIR, possibly influenced by changes in water composition that affect scattering properties. The visible spectrum exhibits weak to moderate correlations, with some variability around the green (~500 nm) and red (~700 nm) regions. These results suggest that pH primarily influences water optical properties in the NIR range, potentially reflecting changes in water chemistry or particulate matter that modulate reflectance.

The turbidity plot exhibits weaker correlations overall compared to nitrates and pH (Figure 8c). However, negative correlations are observed in the NIR region (800–900 nm), with a minimum around 870 nm (−0.55). This negative relationship aligns with expectations, as increased turbidity typically leads to reduced NIR reflectance due to enhanced absorption and scattering effects. The visible spectrum shows weak to moderate variability, with isolated peaks around 550 nm, likely corresponding to green reflectance influenced by suspended particulates or algal content.

The correlation plot for chlorophyll-a lacks significant peaks, suggesting weak or inconsistent relationships across the spectral range (Figure 8d). This result may indicate that chlorophyll-a concentrations have a diffuse or indirect influence on reflectance patterns, or that other factors, such as turbidity or DOM, dominate the optical signal. However, minor variability is noted in the green and red regions (~500–700 nm), consistent with known chlorophyll absorption bands, albeit with weak correlations.

4.3.2. Campaign 2

The nitrate correlation plot demonstrates notable peaks in both positive and negative directions. High positive correlations are observed around 500 nm (~0.83) and 700 nm (~0.85), aligning with the green and red spectral regions (Figure 9a). This suggests that increased nitrate concentrations might enhance reflectance due to interactions with organic or particulate matter, which could amplify scattering in these regions. Conversely, strong negative correlations in the NIR region (~800–900 nm, reaching −0.88) indicate that nitrate-enriched conditions suppress NIR reflectance, likely due to increased water absorption or turbidity effects that reduce backscatter in this spectral range.

The pH correlation plot shows relatively moderate and consistent negative correlations throughout the spectral range, with the most pronounced dips observed around 500 nm (−0.55) and 700 nm (−0.51) (Figure 9b). These correlations indicate that pH variations have a subtle but persistent effect on reflectance, with higher pH potentially reducing optical backscatter in the visible spectrum. The lack of sharp peaks suggests that pH may not directly dominate spectral responses but rather subtly modulates reflectance through its influence on water chemistry or particulate dynamics.

For turbidity, the correlation plot reveals weaker overall relationships compared to nitrates and pH (Figure 9c). However, notable trends include a negative correlation peak near 500 nm (~−0.51) and a positive correlation in the NIR range (~800–900 nm, peaking at 0.54). These findings suggest that turbidity has a complex influence on spectral behavior: it likely suppresses reflectance in the visible spectrum due to increased absorption and enhances it in the NIR region, where scattering effects might become more prominent in turbid waters.

The chlorophyll-a plot displays low correlations across the spectrum, with minor positive peaks around 830 nm (0.52) and negative peaks around 795 nm (−0.63) (Figure 9d). These results suggest a weak but observable relationship between chlorophyll-a and reflectance, potentially influenced by its absorption characteristics in the visible spectrum and secondary scattering effects in the NIR. The modest peaks highlight the need for more targeted spectral analysis to better isolate the chlorophyll-a signal from confounding factors.

4.3.3. Campaign 3

The nitrate correlation plot shows strong positive correlations at wavelengths around 500 nm (~0.64) and 850 nm (~0.66) (Figure 10a). These peaks in the green and NIR regions suggest a link between nitrate levels and increased reflectance in these spectral bands, possibly due to nitrate’s association with particulate or DOM that influences light scattering. Negative correlations are evident in the NIR region (800–900 nm), with a notable dip around 900 nm (−0.52). This indicates that higher nitrate concentrations might reduce NIR reflectance due to enhanced absorption effects or increased water turbidity.

The pH correlation plot highlights consistently negative correlations across the spectrum, with minima around 500 nm (−0.56), 700 nm (−0.60), and 800 nm (~−0.51) (Figure 10b). These results suggest that higher pH levels might be linked to reduced reflectance, particularly in the visible and NIR regions. The relationship is subtle but may reflect changes in water chemistry or particulate interactions that modulate reflectance properties.

For turbidity, the correlation plot displays weaker relationships overall, with minor positive peaks in the green region (~500–600 nm) and the NIR (~800–900 nm) (Figure 10c). The highest correlation is observed around 900 nm (<0.50), indicating that turbidity has a slight influence on NIR reflectance, likely due to increased scattering by suspended particles. The visible spectrum correlations are less pronounced, suggesting a complex interplay between turbidity and reflectance.

The chlorophyll-a plot shows weak correlations across the spectrum, with minor variability in the green and NIR regions (Figure 10d). Positive correlations around 500 nm and 900 nm (both <0.50) suggest a faint link between chlorophyll-a and reflectance in these bands. However, the low magnitude of these correlations indicates that chlorophyll-a’s spectral signature is overshadowed by other factors, such as turbidity or DOM.

4.4. Random Forest Regression

4.4.1. Campaign 1

The feature importance analysis for nitrate concentrations revealed that the most influential spectral regions were concentrated around 800 nm in the NIR region, with a peak importance value of 0.175 (Figure 11). Despite this, the model exhibited poor predictive performance, as indicated by a CV score of 0.08, RMSE of 1.08, R² of −0.01, and RPD of 0.99. The negative R² suggests that the model does not generalize well and provides little improvement over a mean predictor.

For pH, the most relevant spectral features were identified between 750 nm and 850 nm, with the highest contribution occurring at approximately 800 nm. These wavelengths may be indirectly linked to DOM or ionized compounds affecting water reflectance. While the R² of 0.51 indicates that the model explains about 51% of the variance, its CV score of 0.06, RMSE of 0.46, and RPD of 1.43 suggest that, although it outperforms the models for other parameters, its predictive capacity remains moderate.

The prediction of turbidity was largely influenced by wavelengths in both the visible (450–550 nm) and NIR (750–850 nm) regions, aligning with known optical responses related to particle scattering (Figure 11). However, the model’s CV score of −0.42, RMSE of 1.14, R² of −0.26, and RPD of 0.89 indicate that it performed poorly, likely due to the high variability in suspended particle composition, size, and concentration, which reduces spectral consistency.

For chlorophyll-a, key spectral contributions were observed in the 450–550 nm and 680 nm regions, which correspond to known chlorophyll absorption peaks. However, the model yielded weak predictive results, with a CV score of −0.42, RMSE of 1.03, R² of −0.27, and RPD of 0.88, suggesting that spectral interference from other water constituents or relatively low chlorophyll-a concentrations may have hindered model performance.

4.4.2. Campaign 2

For nitrate concentrations, the most influential spectral bands were distributed across the visible and near-infrared regions, particularly around 550 nm, 600 nm, and 700 nm, with a maximum feature importance of 0.08 (Figure 12). Despite the presence of notable wavelengths, the model failed to provide reliable predictions, as indicated by its CV score of −0.30, RMSE of 0.74, R² of −0.57, and RPD of 0.79, reinforcing the difficulty of using hyperspectral data to detect nitrates in mixed waters.

The feature importance analysis for pH revealed spectral peaks in both the visible (550–600 nm) and NIR (750–850 nm) ranges. These findings may suggest indirect relationships with organic matter or other optically active substances affecting reflectance. However, the model’s CV score of −0.77, RMSE of 0.47, and RPD of 1.10 indicate weak predictive capacity, while the R² of 0.18 confirms that the model explains only a small fraction of the pH variability.

For turbidity, the model identified dominant wavelengths between 450 nm and 600 nm, with secondary peaks in the 750–850 nm range (Figure 12). While these spectral responses align with known optical scattering and absorption properties of suspended particles, the CV score of −0.35, RMSE of 1.51, R² of −0.52, and RPD of 0.81 indicate poor predictive performance, suggesting that the heterogeneous composition of suspended matter limits the reliability of hyperspectral models for turbidity estimation.

The chlorophyll-a model identified significant spectral contributions around 500 nm and 680 nm, consistent with expected chlorophyll absorption bands. However, despite this alignment, the CV score of −0.51, RMSE of 0.41, R² of −1.30, and RPD of 0.66 highlight severe predictive limitations, suggesting that low chlorophyll-a concentrations or spectral overlap with other constituents significantly reduced the model’s ability to extract meaningful relationships.

4.4.3. Campaign 3

The feature importance analysis for nitrates indicated strong spectral contributions around 600 nm and 750–800 nm, with a maximum feature importance of 0.06 (Figure 13). While these wavelengths were identified as relevant, the model’s CV score of −0.24, RMSE of 1.27, R² of 0.02, and RPD of 1.01 suggest that it marginally outperformed a simple mean predictor but lacked strong predictive accuracy.

For pH, important wavelengths were observed across the visible and NIR regions, particularly near 500 nm, 650 nm, and 750–850 nm (Figure 13). These may be indirectly associated with organic matter or mineral content variations affecting water reflectance. Despite these identified spectral contributions, the model’s CV score of −1.39, RMSE of 0.68, R² of 0.01, and RPD of 1.01 indicate poor predictive reliability, explaining only 1% of the variance.

The feature importance results for turbidity highlighted key wavelengths in the 450–600 nm and 750–850 nm ranges, corresponding to scattering and absorption mechanisms in water. However, the model performed worse than a simple mean predictor, with a CV score of −0.33, RMSE of 0.72, R² of −0.50, and RPD of 0.81, indicating that the spectral variability introduced by suspended solids was not effectively captured.

For chlorophyll-a, the most relevant spectral regions included 450–550 nm and 680 nm, aligning with known chlorophyll absorption features. Nevertheless, the model exhibited limited predictive power, with a CV score of −0.95, RMSE of 0.71, R² of 0.21, and RPD of 1.13, suggesting that although its performance was slightly better than random chance, it remained insufficient for accurate chlorophyll-a estimation under field conditions.

5. Discussion

5.1. Campaign 1

Campaign 1 represented a relatively stable period in terms of water-quality parameters, with low variability in nitrate concentrations (mean = 1.51 mg/L, Figure 3a) and stable pH values near neutral (mean = 7.15, Figure 3b). The observed stability in nitrate concentrations aligns with findings that suggest consistent baseline levels in agricultural runoff can lead to predictable nitrate distributions [26] (Balachandran et al., 2020). The pH’s consistency suggests it may serve as a robust indicator of water quality in this system, with minimal susceptibility to temporal or episodic disturbances as found in another research [27] (Kgopa et al., 2018). Turbidity was the lowest among the three campaigns, with minimal variability (mean = 28.9 NTU, Figure 3c), suggesting reduced particulate input into the canal system. In contrast, chlorophyll-a concentrations showed considerable variation (mean = 249.03 mg/m³, Figure 3d), with a broad IQR indicative of localized algal blooms.

Spectral reflectance patterns (Figure 5) exhibited a green peak around 500–600 nm, which is consistent with chlorophyll absorption features [28] (H. Zhang et al., 2024). The decline in reflectance beyond 600 nm and its stabilization around 5% in the NIR region suggests that water absorption dominated over scattering effects, aligning with optical behavior expected in aquatic environments [29,30] (El-Hendawy et al., 2017; Gad et al., 2020). The relatively stable reflectance trends indicate homogeneity in water composition, likely driven by the canal’s controlled hydrological conditions.

The correlation analysis (Figure 8) indicated that nitrate concentrations exhibited moderate positive correlations in the green (~500 nm) and red (~700 nm) spectral regions but strong negative correlations in the NIR (~800–900 nm), likely due to interactions with DOM [31] (Wu et al., 2022). pH showed positive correlations in the NIR (~830 nm), potentially linked to indirect effects on scattering properties, whereas turbidity correlations were weaker, with negative associations in the NIR, suggesting that particulate matter absorption dominated [32] (Yan et al., 2023). Chlorophyll-a showed weak and inconsistent correlations across all wavelengths, highlighting challenges in detecting its spectral signature due to interference from turbidity and DOM.

The RF model (Figure 11) identified 800 nm as the most relevant spectral region for nitrate estimation, but the predictive performance was poor (R² = −0.01, CV = 0.08). For pH, wavelengths between 750 and 850 nm were identified as important, and the model achieved the highest predictive performance among all parameters (R² = 0.51, CV = 0.06). However, models for turbidity and chlorophyll-a were weak, with negative R² values (−0.26 and −0.27, respectively), indicating that spectral variability introduced by particle heterogeneity and low chlorophyll-a concentrations hindered predictive capacity [33] (X. Wang et al., 2017).

5.2. Campaign 2

Campaign 2 exhibited the highest variability in nitrate concentrations (mean = 1.56 mg/L, Figure 3a), with broader IQRs and outliers exceeding 2.5 mg/L, suggesting episodic inputs from agricultural runoff and seasonal changes [34] (Ma et al., 2023). The turbidity levels were the highest among all campaigns (mean = 71.63 NTU, Figure 3c), likely driven by sediment resuspension or increased suspended solids from wastewater inflows. This elevated turbidity may be attributed to increased particulate matter or sediment resuspension, potentially driven by localized disturbances, seasonal runoff, or anthropogenic activities. Research has demonstrated that turbidity levels can vary significantly due to sediment resuspension and runoff events [35] (Goyens et al., 2022). Chlorophyll-a concentrations were the lowest (mean = 97.62 mg/m³, Figure 3d), possibly due to unfavorable environmental conditions or nutrient limitations [36] (Khadr et al., 2021).

Reflectance values in Campaign 2 were the highest across all campaigns, particularly in the NIR region (10–15%, Figure 6). This elevated reflectance is likely linked to increased particulate scattering, a phenomenon commonly observed in turbid waters [36] (Khadr et al., 2021). The distinct reflectance peak at 550 nm suggests the presence of suspended particles influencing optical properties [37] (Guo et al., 2020).

Correlation analysis (Figure 9) indicated that nitrates showed strong positive correlations at 500 nm (~0.83) and 700 nm (~0.85), suggesting that nitrate-enriched conditions influenced scattering effects. However, strong negative correlations in the NIR (~800–900 nm, reaching −0.88) suggested increased absorption due to associated increases in turbidity. These results suggest that nitrate concentrations may influence reflectance indirectly through interactions with DOM or suspended particulates that enhance scattering in the visible range. The negative correlations in the NIR region align with increased water absorption, which may be exacerbated under nitrate-enriched conditions due to associated increases in turbidity or dissolved matter. Previous research has similarly demonstrated that nitrate levels can significantly affect spectral reflectance patterns [31,38] (Gan et al., 2020; Wu et al., 2022). The pH correlations were more diffuse, with consistent negative trends across the spectrum, while turbidity exhibited weak but expected patterns, including negative correlations in the visible spectrum (~500 nm) and positive correlations in the NIR. Chlorophyll-a displayed only minor correlations (~0.52 at 830 nm), reinforcing the difficulty of detecting this parameter in mixed water systems.

The RF model (Figure 12) identified multiple spectral regions (550–600 nm and 750–850 nm) as relevant for nitrate and pH estimation, but the predictive performance remained weak (nitrate R² = −0.57, pH R² = 0.18). The turbidity model identified peaks in the 450–600 nm range, consistent with known scattering mechanisms, but performed poorly (R² = −0.52). The chlorophyll-a model was the least effective (R² = −1.30), likely due to low concentrations and overlapping optical signals, confirming the limitations of hyperspectral reflectance in highly complex wastewater conditions [39] (Alqarawy et al., 2022).

5.3. Campaign 3

Campaign 3 exhibited the lowest variability in nitrate concentrations (mean = 1.28 mg/L, Figure 3a) and stable pH values (mean = 7.21, Figure 3b), suggesting minimal external influences on these parameters. Turbidity levels (mean = 42.38 NTU, Figure 3c) were lower than in Campaign 2 but exhibited moderate variability, likely reflecting seasonal changes in flow conditions. Chlorophyll-a concentrations were the highest among all campaigns (mean = 324.05 mg/m³, Figure 3d), with an extensive IQR and outliers exceeding 400 mg/m³, suggesting episodic algal blooms driven by nutrient inputs. Studies have shown that elevated nutrient levels often correlate with increased algal activity, particularly in systems experiencing episodic nutrient surges [36] (Khadr et al., 2021).

Reflectance in Campaign 3 was the lowest across the three campaigns (median = 2.79%, Figure 7), with reduced NIR reflectance (2–3%), indicating minimal scattering from suspended solids. The subdued green peak (~550 nm) and the steep decline in the NIR region are consistent with lower turbidity and increased water absorption [6,39] (Alqarawy et al., 2022; Yang et al., 2022). The spectral stability suggests a relatively uniform water composition, possibly due to reduced external disturbances.

Correlation analysis (Figure 10) revealed moderate nitrate correlations at 500 nm (~0.64) and 850 nm (~0.66), indicating potential interactions with DOM [31] (Wu et al., 2022). However, strong negative correlations in the NIR (~900 nm, −0.52) suggested enhanced absorption effects. The pH correlation patterns were consistent with those of previous campaigns, with negative correlations near 500 nm (−0.56) and 700 nm (−0.60), indicating indirect influences through water chemistry interactions [40] (Saad et al., 2022). The strongest correlations in the red and NIR regions indicate that pH variations may influence spectral behavior through interactions with dissolved carbonates, organic matter, or particulates that alter optical backscatter. However, the absence of sharp peaks suggests that pH does not exert a dominant influence on hyperspectral reflectance and instead acts as a general modulator of water optical properties [40,41] (Saad et al., 2022; D. Zhang et al., 2022). Turbidity exhibited weak positive correlations in the NIR (~900 nm, <0.50), suggesting some influence on backscattering, while chlorophyll-a correlations remained low across the spectrum.

The RF model (Figure 13) identified spectral contributions around 600 nm and 750–800 nm for nitrate estimation, but the model performance remained weak (R² = 0.02, CV = −0.24). pH feature importance was distributed across 500–850 nm, yet the model explained virtually no variance (R² = 0.01, CV = −1.39). The turbidity model performed the worst (R² = −0.50), suggesting that suspended particle variability was not effectively captured [33] (X. Wang et al., 2017). The chlorophyll-a model showed slightly better performance (R² = 0.21), but its low predictive capacity (CV = −0.95) underscores the limitations of using hyperspectral data for chlorophyll-a retrieval in mixed water environments [41] (D. Zhang et al., 2022).

5.4. Chlorophyll-a

All campaigns reported extreme chlorophyll-a levels, substantially surpassing conventional concentrations observed in natural water bodies. For instance, chlorophyll-a concentrations ranging from 0.7 to 2.6 mg/m³ were found in the Vargem das Flores reservoir and 12 to 18 mg/m³ in the Lagoa Santa lake [42] (Pinto et al., 2001). The extremely high values in this study, particularly the maximum values exceeding 400–600 mg/m³, strongly suggest hypereutrophic conditions consistent with substantial nutrient enrichment from served waters in tropical environments [43] (Carvalho et al., 2022). Such elevated chlorophyll-a levels are often linked to excessive N and P inputs, which drive the proliferation of microalgae in wastewater-influenced systems. This is further supported by research in the Sembrong Dam, Malaysia, where chlorophyll-a concentrations averaging 175.9 µg/L classified the water body as hypereutrophic, highlighting the severe impact of nutrient pollution from surrounding agricultural activities [44] (Wellson et al., 2016).

Similar elevated chlorophyll-a patterns have been documented across diverse wastewater-influenced systems. A comparable trend was observed in microalgae cultivation using tofu wastewater, where chlorophyll-a concentrations reached 6.38 mg/L, demonstrating that nutrient-rich effluents can sustain excessive algal growth [45] (Taufikurahman et al., 2024). Research on urban canal systems has highlighted that chlorophyll-a variability is strongly influenced by short-term weather fluctuations and anthropogenic activities, with higher concentrations often recorded in stagnant water segments and slow-flowing canals [46] (Zhou et al., 2021). Further supporting these findings, microalgal blooms in wastewater-impacted environments have been associated with the simultaneous accumulation of toxic metals, which may act as additional stressors or selective agents for specific algal species, altering biomass composition and productivity [47] (Muthuraman et al., 2023). The presence of dominant microalgae species such as Botryococcus in hypereutrophic environments suggests that excessive algal growth not only disrupts water quality but may also interfere with water treatment processes by clogging filters and increasing treatment costs [44] (Wellson et al., 2016).

Recent research on lakes recharged by reclaimed water found that NO₃⁻-N is the primary nutrient driver of eutrophication, particularly under high-temperature conditions, where it significantly correlates with chlorophyll-a concentration [48] (C. Wang et al., 2024). This study emphasized that nitrate input from reclaimed water, rather than total N alone, plays a critical role in phytoplankton proliferation, suggesting that wastewater-irrigated agricultural canals are highly susceptible to seasonal algal blooms driven by N dynamics. These findings highlight the need for improved nutrient management strategies in wastewater-reuse irrigation systems. Moreover, the variability in correlation strength across campaigns reflects the dynamic nature of turbidity and its dependence on particle composition, size, and concentration. The relatively weaker correlations compared to nitrates suggest that while turbidity influences reflectance, its effects are more complex and potentially confounded by other water-quality parameters.

5.5. pH Levels

The ability to predict pH using infrared spectral data in our agricultural irrigation system, despite its moderate correlation strength (r ≈ 0.51), can be attributed to several indirect mechanisms related to suspended materials in this complex mixed water environment. While pH itself does not directly absorb light in the visible to near-infrared spectrum, it significantly influences the optical properties of suspended particles and dissolved constituents that do interact with electromagnetic radiation. In agricultural irrigation canals receiving served water, pH often correlates strongly with the composition and concentration of suspended materials including clay minerals, metal oxides/hydroxides, and organic matter aggregates. These materials exhibit pH-dependent surface charges that alter their aggregation behavior, particle size distribution, and consequently, their light scattering and absorption properties [2] (Gholizadeh et al., 2016). For instance, aluminum and iron hydroxides undergo significant structural and optical changes at varying pH levels, affecting their reflectance across the NIR region (450–900 nm used in our study). Additionally, pH-induced changes in microbial communities and algal composition—which were abundant in our samples as evidenced by high chlorophyll-a levels—create distinctive spectral signatures detectable in the 700–850 nm range. The interdependent relationship between pH and suspended material characteristics is particularly relevant in our irrigation system, where mixing of wastewater effluent with varying pH levels creates dynamic equilibria with suspended solids, resulting in detectable spectral patterns. These findings align with recent work demonstrated that indirect estimation of optically inactive parameters was possible in complex water systems through their covariation with optically active constituents [49,50,51,52] (Ahmed et al., 2022; Fu et al., 2024; Niu et al., 2021; Raheli et al., 2024).

6. Conclusions

This study evaluated the potential of in situ HRS for monitoring key water-quality parameters—nitrates, pH, turbidity, and chlorophyll-a—within an agricultural mixed-water irrigation canal across three field campaigns. While model performance was suboptimal, the results provide valuable insights into spectral relationships with water quality in complex mixed-water systems.

The descriptive analysis revealed that while nitrates and pH exhibited relative stability across campaigns, turbidity and chlorophyll-a showed significant variability, highlighting the dynamic nature of the studied system. The highest turbidity levels in Campaign 2 were associated with increased sediment resuspension, whereas Campaign 3 recorded the highest chlorophyll-a concentrations, indicating episodic algal blooms. Spectral reflectance trends confirmed that the NIR region (750–900 nm) was critical in distinguishing changes in water composition, particularly for parameters influenced by suspended solids and DOM.

The correlation analysis identified nitrates as the most spectrally responsive parameter, with notable relationships in the green (~500 nm), red (~700 nm), and NIR (~800–900 nm) regions. However, weaker and inconsistent correlations were observed for pH, turbidity, and chlorophyll-a, suggesting that their interactions with hyperspectral signals were influenced by external environmental factors or co-occurring constituents.

The RF regression analysis further demonstrated the complexity of modeling mixed water quality with hyperspectral data. The feature importance analysis highlighted relevant spectral regions, particularly in the blue-green (450–550 nm), red (680–700 nm), and NIR (750–850 nm) ranges, but the predictive performance remained poor across all campaigns. Consistently low cross-validation scores and negative R² values suggest that indirect relationships and spectral interference limit the reliability of hyperspectral models for these parameters in mixed water conditions.

Despite these limitations, this study provides a baseline understanding of spectral trends associated with mixed-water quality in tropical agricultural systems. The identification of key spectral regions supports the potential refinement of hyperspectral techniques for water monitoring. Future research should focus on expanding the spatial and temporal scope of sampling, incorporating additional water-quality parameters, and validating findings across diverse agricultural settings to support sustainable water reuse in irrigation. These efforts will be essential for optimizing hyperspectral approaches in addressing the complexities of mixed water systems.