Analysis of Reservoir Water Quality by Smartphone Color Image Analysis: A Case Study of Three Reservoirs in Taiwan

Featured Application

The proposed AI-assisted Water Health Index (WHI) system enables rapid, low-cost, and participatory monitoring of reservoir water quality using only smartphone imagery. This approach can be applied for early detection of turbidity increase, algal blooms, or nutrient pollution in drinking water and agricultural reservoirs, supporting real-time management decisions in resource-constrained regions. This application concept is supported by recent findings on mobile image-based water-quality diagnostics.

Abstract

This work investigates smartphone-based image processing for monitoring reservoir water quality, driven by the necessity for accessible and economical environmental evaluation techniques. The objective is to examine the correlation between water color and essential water quality parameters, including turbidity, total phosphorus, and chlorophyll-a, utilizing basic, readily accessible technology. The concept entails taking water photos from three principal reservoirs in Taiwan—Shimen, Liyutan, and Hushan—utilizing a smartphone camera, succeeded by sophisticated image processing algorithms, encompassing RGB color space analysis and ripple filtering. The findings indicate strong correlations between the G/R ratio obtained from the photos and conventional water quality parameters, particularly turbidity and chlorophyll-a. The correlation analysis yielded R² = 0.72 (p < 0.01) for turbidity and R² = 0.68 (p < 0.05) for chlorophyll-a, confirming the statistical significance of the results.

1. Introduction

Monitoring water quality is essential for efficiently managing and preserving water resources, especially in reservoirs that are key drinking water sources and agriculture. Conventional water quality assessment methods frequently require laborious and expensive laboratory tests, necessitating the development of more efficient and economical monitoring systems [1,2].

In recent years, utilizing remote sensing and image processing methodologies for water quality evaluation has garnered considerable interest. Satellite images and aerial photography have been employed to assess multiple water quality measures, such as turbidity, chlorophyll-a concentration, and suspended solids [3,4]. Nonetheless, these techniques frequently necessitate costly apparatus and specialized knowledge, constraining their broad implementation. Remote sensing systems often require specialized knowledge in image calibration, spectral correction, and data interpretation, which limits their accessibility for routine or community-based monitoring applications.

1.1. Water Quality Telemetry

Remote sensing technology has developed into a potent instrument for water quality evaluation, employing the interplay between sunlight and aquatic environments to analyze numerous parameters. Visible light sensors can identify alterations in water color or turbidity, whereas near-infrared sensors can assess characteristics such as absorption, reflection, and scattering [5]. Smartphone-based sensing can be regarded as a natural extension of conventional remote sensing principles, transferring the concept of reflectance-based observation from satellite or aerial platforms to handheld devices equipped with RGB sensors. Recent research has demonstrated significant relationships between spectral data and water quality indicators, such as chlorophyll-a, clarity, turbidity, and total phosphorus [3,6]. The amalgamation of remote sensing and machine learning methodologies has significantly improved the precision and efficacy of water quality [1]. Devised an innovative method that integrates remote sensing data with deep learning algorithms to assess water quality indicators in inland waters, exhibiting enhanced accuracy relative to conventional techniques.

1.2. The Correlation Between Water Color and Water Quality

The color of water acts as an intuitive signal of its quality, affected by numerous substances inside it. Recent research has shown that water color can serve as an indicator for multiple water quality measures, such as dissolved organic matter, chlorophyll-a, and suspended sediments [4,7].

Advancements in hyperspectral imaging have enhanced our capacity to identify nuanced alterations in water color and associate them with certain water quality metrics. Employed hyperspectral imaging to create an innovative technique for assessing chlorophyll-a content in inland waters with good accuracy in various aquatic settings [6].

1.3. Algal Analysis

Algal analysis is essential for evaluating water quality, as algal blooms can profoundly affect aquatic ecosystems and human health. Recent improvements in fluorescence-based technologies have facilitated expedited and automated identification and classification of algae [8,9].

Combining machine learning with spectrum imaging has transformed algal detection and categorization. Devised a convolutional neural network methodology for the automated identification and quantification of harmful algal blooms utilizing multispectral satellite images, exhibiting significant accuracy and efficiency in bloom detection [9].

1.4. RGB Color Space and Image Processing

The RGB color space, founded on the trichromatic characteristics of human vision, is extensively utilized in digital picture analysis for environmental monitoring. Recent research has investigated the feasibility of utilizing RGB values from digital photographs to assess several water quality metrics [10,11]. Improvements in image processing methodologies, encompassing artificial intelligence and computer vision, have significantly augmented the efficacy of picture-based water quality evaluation [12,13].

Affordable, consumer-level imaging technologies have created new opportunities for extensive water quality monitoring. Illustrated the efficacy of smartphone cameras in assessing turbidity in urban water bodies, emphasizing the accessibility and cost-effectiveness of this method [14].

Notwithstanding these developments, a gap persists in the literature about utilizing readily accessible, consumer-grade equipment for thorough reservoir water quality monitoring. The extensive accessibility of high-quality smartphone cameras offers a chance to provide a more accessible and economical approach to evaluating water quality. This work seeks to fill this gap by exploring the feasibility of employing smartphone photography and image analysis methods to assess reservoirs’ water quality. We examine the relationships between RGB values obtained from smartphone photos and several water quality measures assessed using conventional methods. Our research aims are as follows:

• To formulate and authenticate a methodology for evaluating water quality through smartphone photography and image analysis.
• To examine the relationships between image-derived color metrics, specifically the G/R ratio, and essential water quality measures in three Taiwanese reservoirs.
• To assess the advantages and constraints of this method as an adjunct to traditional water quality monitoring techniques.

This study aims to enhance the accessibility and affordability of water quality monitoring instruments by utilizing readily available technologies and establishing a straightforward yet efficient image analysis procedure. The results of this research may substantially influence the frequency and spatial extent of water quality evaluations, especially in areas with constrained resources for environmental monitoring. In summary, this study introduces a novel, low-cost, and easily replicable method for reservoir water-quality monitoring that integrates smartphone-based RGB image analysis, ripple filtering, and an AI-assisted Water Health Index (WHI). The novelty of this work lies in combining consumer-grade smartphone imaging with AI-driven colorimetric interpretation, marking the first application of such an integrated system for Taiwan’s reservoir management. Recent smartphone-based studies have reported comparable accuracy in estimating turbidity and chlorophyll-a using simple imaging systems [13,14]. These examples demonstrate the feasibility of low-cost handheld devices for rapid and accessible environmental monitoring. Several studies have demonstrated the potential of hyperspectral and satellite-based imaging for estimating key water-quality parameters, providing a foundation for image-based analysis approaches [15,16].

2. Materials and Methods

2.1. Sampling Points and Methods

Our study thoroughly investigated water quality at three important reservoirs: Shimen Reservoir in Taoyuan City, Liyutan Reservoir in Miaoli County, and Hushan Reservoir in Yunlin County. We used a smartphone (iPhone 12, Apple Inc., Cupertino, CA, USA) equipped with a circular polarizer (K&F Concept, Guangzhou, China) and a 18% gray card (X-Rite, Grand Rapids, MI, USA) for image sampling. The study spanned a period of eight months, from February to September this period covers both the dry and rainy seasons in Taiwan, allowing the dataset to reflect seasonal variations in turbidity, transparency, and algal growth. This inclusion strengthens the representativeness of the monitoring data across different hydrological conditions. A comprehensive dataset of 32 sampling events was obtained by conducting monthly samplings at two crucial locations in each reservoir: the water inflow and the water intake. These samplings were carried out at both surface and mid-levels, ensuring a thorough analysis of the reservoir conditions. We utilized sophisticated gear to evaluate water quality, using portable meters to measure several parameters such as pH, conductivity, dissolved oxygen, turbidity, and chlorophyll levels. Surface samples were obtained from a depth of 0.5 m below the water surface, while mid-level samples were collected from a depth of 10 Surface samples were obtained from a depth of 0.5 m below the water surface, while mid-level samples were collected from a depth of 10 m. Pre-rinsed sampling tubes were used to maintain the integrity of the samples. Deeper water samples were collected using a water sampler (Wildco Model 1120-G43, Wildlife Supply Company, Yulee, FL, USA), while mid-level samples were acquired using a depth sampler activated by a nylon rope mechanism. Water transparency was measured using a Secchi disk (model 20 cm, Hach Company, Loveland, CO, USA) by determining the depth at which it became invisible. We used a smartphone (iPhone 12, Apple Inc., Cupertino, CA, USA) equipped with a circular polarizer (K&F Concept, Guangzhou, China) and an 18% gray card (X-Rite, Grand Rapids, MI, USA) for image sampling. The choice of this smartphone model was therefore based on its technical stability, color fidelity, and reproducible imaging performance—not on availability. This combination served to minimize polarized light reflections and guarantee optimal image clarity. The gray card ensures precise white balance by accounting for aberrations caused by ambient lighting. To reduce disturbances caused by boat movements, the phone was positioned at an angle of 30–45 degrees relative to the water’s surface. The polarizer was calibrated to minimize polarization effects, and the white balance was meticulously adjusted using the gray card. The phone’s automated metering mode ensured the ideal level of exposure, collecting five photographs at each sample point, which were then organized into three sets to provide thorough coverage. The methodology is demonstrated in Figure 1. The study’s substantial addition to the field of environmental monitoring is based on our advanced technology utilization and thorough sampling method, which enable a full understanding of water quality in various environmental circumstances.

2.2. Water Sample Testing Items and Methods

Our study utilizes in situ measurements and laboratory investigations to evaluate water quality comprehensively. We assess various factors, including pH, temperature, electrical conductivity (EC), dissolved oxygen (DO), total phosphorus (TP), total nitrogen (TN), total organic carbon (TOC), suspended solids (SS), clarity, turbidity, and chlorophyll a. In addition, we quantify and identify the algal species found in the water samples.

We utilized portable meters (Hach HQ40D, Hach Company, Loveland, CO, USA) to conduct on-site tests for pH, electrical conductivity (EC), dissolved oxygen (DO), turbidity, and chlorophyll levels. We employed cylinders and depth samplers (Wildco Model 1120-G43, Wildlife Supply Company, Yulee, FL, USA) to ensure accurate collection of surface samples taken from a depth of 0.5 m below the water surface and mid-level samples obtained from a depth of 10 m. Water transparency was quantified using a Secchi disk (model 20 cm, Hach Company, Loveland, CO, USA), which was submerged in water until it became invisible, allowing for the measurement of water clarity based on the depth at which it disappears.

We conducted meticulous examinations within the laboratory to validate the data collected on-site and to quantify supplementary variables. The bbe-Moldaenke FluoroProbe III (bbe Moldaenke GmbH, Schwentinental, Germany) is an advanced fluorometric device used for algal examination. It employs six separate LED wavelengths to assess algal fluorescence and chlorophyll concentration. This analyzer provides real-time monitoring and visualization of chlorophyll levels, algal counts, depth, and water temperature. The recorded data can be analyzed in detail on a connected laptop. This comprehensive strategy guarantees precise and dependable evaluations of water quality, offering significant observations regarding the ecological well-being of the reservoirs.

2.3. Image Processing Technology

This section focuses on using sophisticated image processing techniques to improve the accuracy of water quality evaluations through colorimetric analysis. Using Python (version 3.10.12, Python Software Foundation, Wilmington, DE, USA) and the OpenCV package (version 4.9.0, Intel Corporation, Santa Clara, CA, USA), we extracted each image’s Region of Interest (ROI), specifically targeting the middle area filtered by a circular polarizer. This region is essential for precise analysis, enabling us to separate the image into its red (R), green (G), and blue (B) channels. Individual channels are processed independently, yielding grayscale images with intricate color information. Ripple filtering is crucial in our method, eliminating distortions produced by surface ripples or unwanted chemicals. This method entails establishing customized threshold values to distinguish and remove the unwanted oscillations, substituting them with the mean grayscale value of the authentic water color pixels. As seen in Figure 2, the depicted method guarantees that the final image faithfully portrays the water’s inherent condition.

Figure 2 delineates the complete procedure for picture sampling and processing in the evaluation of water quality. The procedure commences with the initial picture acquisition (B1) and advances through multiple processing stages, encompassing type segmentation, low and high threshold filtering (B2, B6), binarization (B3), and value truncation (B7). These stages result in the production of two final outputs: Diagram Final and Diagram Solid (B8). The Diagram Final depicts a comprehensive, colorized rendition of the processed image, complete with annotated RGB averages, whilst the Diagram Solid displays a streamlined solid color depiction of the average values. This systematic approach improves the accuracy of water color analysis, offering a dependable way for assessing ambient water quality. By meticulously classifying and storing these enhanced photos, along with their RGB values and ratios, the system provides a detailed monthly assessment of water quality at each sampling site. This sophisticated image processing method exhibits considerable promise for precise and swift water quality evaluation, enhancing data accuracy and consistency while facilitating further applications in environmental monitoring.

2.4. Correlation Analysis

Correlation analysis is a statistical method employed to evaluate the associations between two or more sets of variables, revealing how they co-vary within a range of values. In statistics, the relationship between variables is measured using positive, negative, or numerical values, indicating their link’s strength and direction. Two often used approaches are Pearson and Spearman correlation analyses, which use a correlation coefficient ranging from −1 to 1 to quantify the strength of the association. A coefficient close to 1 indicates a strong positive correlation, suggesting that if one parameter increases, the other also increases. On the other hand, a coefficient approaching −1 implies a significant negative correlation, meaning that an increase in one variable is likely to result in a reduction in the other variable. In addition to the G/R ratio, additional spectral ratios such as B/R and G/B were analyzed to assess their comparative response to multiple water-quality indicators. This multi-ratio approach improves robustness and helps validate colorimetric relationships across different spectral components.

2.5. Image Calibration Procedure

We instituted a stringent calibration method to maintain consistency in reflectance readings across varying time periods and environmental conditions. This procedure involves utilizing a standard 18% gray card and a color checker chart, both captured under same lighting circumstances as the water surface, thus ensuring constant reference points for color correction. All photographs are taken using manual exposure settings, with precise ISO, aperture, and shutter speed documented for each photograph. A bespoke white balance is established utilizing a gray card for each sample session to accommodate fluctuations in ambient light color temperature. Our post-processing workflow entails capturing images in RAW format, generating a color profile for each sampling session utilizing a gray card and color checker, applying this profile to standardize colors across all images from that session, and converting the standardized images to the sRGB color space for analysis. Furthermore, we save detailed metadata for each photograph, encompassing date, time, GPS locations, weather conditions, and all camera settings, facilitating future retrospective adjustments and ensuring reproducibility. This calibration technique markedly enhances the consistency and reliability of our color readings, facilitating more precise multi-period analyses of water quality.

2.6. AI-Based Water Health Index (WHI) Summarization

To translate multi-parameter predictions into a single interpretable metric, we developed an AI-assisted Water Health Index (WHI). The WHI framework integrates RGB-derived features (e.g., G/R, G/B ratios, luminance statistics) with measured parameters (turbidity, transparency, chlorophyll-a, phosphorus, algal counts) using a multi-task ensemble architecture consisting of Random Forest regressors and Gradient Boosting models. Each model predicts the target parameters simultaneously, and the outcomes are normalized to a 0–100 scale according to ecological thresholds. Weighted aggregation of these sub-scores yields the WHI, which classifies reservoir condition into four categories: Excellent (85–100), Good (70–84), Moderate (50–69), and Poor (<50). Uncertainty is quantified by bootstrap ensembles, providing 95% prediction intervals (PIs). This approach produces a concise water quality summary from a smartphone image in near real-time.

3. Results

This chapter covers the findings of our investigation, specifically examining the efficacy of image processing algorithms and their relationship with water quality measures. We conducted a well-planned study where we collected water samples from three reservoirs in Taiwan for eight months. During each sampling occasion, we took 15 photos. The photos underwent thorough analysis to consider factors that could potentially distort the actual color of the water, such as suspended particles, foam, and natural detritus. Advanced filtering techniques were employed to reduce the impact of ripples, which are recognized as the primary determinant of water color. In addition to ripples, we also considered other factors, such as foam, leaves, and suspended particulates, which improved the precision of our water quality visualization. The efficacy of our image processing was verified by comparing unprocessed and processed image data with water quality parameters from February to July. The correlation between key indicators, such as turbidity, transparency, diatoms, and total algal count, showed significant improvements. This indicates that our processed data is reliable, as demonstrated in

Table 1. Processing Comparison of R² before and after image processing (values with p < 0.05 are marked as significant).

Parameter	Before Filtering (R2)	After Filtering (R2)
Turbidity and R2 of G/R	0.5435	0.5449
Transparency and R2 of G/R	0.6785	0.7219
R2 of diatoms and G/R	0.6025	0.6367
R2 of total cell count and G/R	0.7104	0.7212

. This table compares the R² values before and after picture processing. It clearly shows a significant improvement in the correlation with these parameters, highlighting the effectiveness of our advanced image-processing techniques.

The post-processing photographs were methodically organized by reservoir and month, with selected images showcasing differences in water color. The monthly water color photographs of Hushan Reservoir, as depicted in Figure 3, showcase a range of primary colors, varying from vivid green to dark green. The G/R and G/B ratios influence these colors. This diagram offers a graphical overview of the fluctuations in water color over the year, illustrating the ever-changing environmental circumstances.

Table 2. Image data of Hushan Reservoir from February to August.

Image Number	R	G	B	Ledger	Pick-Up/Inclusion	Gross Income
Serial: H0201	89.00	160.84	106.92	1.81	1.20	1.50
Serial: H0202	92.59	149.59	99.83	1.62	1.08	1.50
No: H0301	113.85	143.76	98.69	1.26	0.87	1.46
Serial: H0302	94.23	144.45	91.58	1.53	0.97	1.58
Number: H0401	90.11	145.38	101.62	1.61	1.13	1.43
Number: H0402	83.49	143.71	102.28	1.72	1.23	1.41
Number: H0501	90.55	129.51	74.63	1.43	0.82	1.74
Ref: H0502	94.98	131.40	78.72	1.38	0.83	1.67
Serial: H0601	95.00	149.12	79.56	1.57	0.84	1.87
Ref: H0602	90.83	144.78	75.31	1.59	0.83	1.92
Serial: H0701	93.03	125.58	78.14	1.35	0.84	1.61
Serial: H0702	78.88	110.11	62.40	1.40	0.79	1.76
Number: H0801	109.01	143.83	84.16	1.32	0.77	1.71
Serial: H0802	102.86	138.39	76.24	1.35	0.74	1.82
Number: H0901	104.48	145.64	70.83	1.39	0.68	2.06
Serial: H0902	77.42	117.61	66.41	1.52	0.86	1.77

presents comprehensive picture data for Hushan Reservoir, illustrating the monthly fluctuations in RGB values and their corresponding ratios. This data is crucial for comprehending the environmental variables that impact water color and quality throughout the year. We conducted a thorough correlation analysis between picture data and water quality measures using Spearman’s rank correlation coefficient. The surface water data showed strong associations, especially with the G/R ratio, highlighting the potential of using image-based monitoring to determine water quality. The correlation analysis conducted for Shimen, Liyutan, and Hushan Reservoirs provided detailed evidence of clear connections between picture data and water quality measures. These findings improve our comprehension of how visual criteria indicate water quality fluctuations in various reservoirs. In summary, our work shows that by applying pre-processing and advanced image processing techniques, the accuracy of using water color photographs to estimate water quality is much enhanced. This establishes a dependable basis for future developments in environmental monitoring.

3.1. Sampling Analysis Results from Shimen Reservoir, Liyutan Reservoir, and Hushan Reservoir

Our study thoroughly analyzed water quality data and photos, uncovering substantial correlations among essential parameters like transparency, turbidity, diatoms, and total algae. The relationships are illustrated in graphs for each reservoir and are further upon below. Figure 4 illustrates the trends in water quality for the Shimen, Liyutan, and Hushan Reservoirs from February to August. Each reservoir is depicted by two graphs: one illustrating transparency and turbidity, and another displaying diatom and total algal numbers.

Figure 4a, depicting Shimen Reservoir, illustrates a distinct negative association between clarity and turbidity from February to August. Transparency progressively diminishes from around 4 m to 2 m, and turbidity escalates from about 4 NTU to 7 NTU. The graph depicting diatoms and total algae demonstrates steady growth, with total algae consistently surpassing diatoms, escalating from approximately 900 cells/mL to 2800 cells/mL for diatoms and from 1200 cells/mL to 3600 cells/mL for total algae.

The Liyutan Reservoir, illustrated in Figure 4b, exhibits a pattern akin to that of Shimen, albeit with certain distinctions. The reduction in clarity and rise in turbidity seem relatively mild. Transparency diminishes from approximately 5 m to 3 m, and turbidity escalates from roughly 3 NTU to 6 NTU. The diatom and total algal counts exhibit a progressive increase, resembling the pattern observed in Shimen, albeit with marginally lower values.

The Hushan Reservoir, depicted in Figure 4c, exhibits a unique pattern. Transparency initiates at an elevated level (about 5 m) and its decline is slower, decreasing to just under 4 m by August. Turbidity exhibits a modest rise from around 4 NTU to 6 NTU; however, the trend is less noticeable in comparison to the other reservoirs. The diatom and total algal counts in Hushan exhibit an upward tendency but at lower overall levels than those observed in Shimen and Liyutan.

Throughout all three reservoirs, a consistent trend of diminishing transparency and escalating turbidity is evident from February to August. This correlates with rising diatom and overall algal counts, indicating a connection between algal proliferation and water clarity. The extent and pace of these alterations differ among the reservoirs, underscoring the intricate interaction of local environmental and operational factors influencing water quality at each site.

This analysis highlights the necessity of evaluating both physical (transparency, turbidity) and biological (algal counts) characteristics for assessing water quality in reservoirs. This also illustrates the considerable variability of these properties among different water bodies, even within the same region and timeframe.

3.2. Correlation Analysis Between Image Data and Experimental Analysis Data

We performed an extensive examination of the correlation between picture data and water quality metrics in several reservoirs. To ensure statistical reliability, significance testing (p < 0.05) and 95% confidence intervals were added for each correlation coefficient, allowing more accurate interpretation of the strength and uncertainty of each relationship. We analyzed these correlations using Spearman’s rank correlation coefficient and illustrated our results in a comprehensive heatmap visualization (Figure 5). All correlation coefficients were accompanied by 95% confidence intervals and significance levels. Statistically significant relationships (p < 0.05) are explicitly indicated in Table 1 and Figure 5 and Figure 6 using asterisks.

Figure 5 depicts the correlation study between image data ratios (G/R, B/R, G/B) and multiple water quality metrics for surface water in all reservoirs. This visualization integrates previously distinct investigations of surface water, mesopelagic water, and algal counts. Our investigation demonstrated significant associations, especially for the G/R ratio. The G/R ratio exhibited a robust positive link with transparency (0.87) and significant negative correlations with turbidity (−0.76), total phosphorus (−0.70), orthophosphate (−0.82), dissolved phosphorus (−0.81), chlorophyll-a (−0.77), diatoms (−0.82), and total algal count (−0.83).

The B/R ratio exhibited limited strong associations with water quality measures, except for a robust positive link with total algal count (0.82). The G/B ratio typically shown weak to moderate negative associations with the majority of factors. These findings highlight the G/R ratio’s potential as a crucial indication for several water quality metrics, especially for surface water. The significant correlations identified indicate that image-based analysis, particularly utilizing the G/R ratio, may serve as an effective instrument for the swift evaluation of water quality in reservoir settings. The heatmap visualization adeptly illustrates the intricate correlations between picture data and diverse water quality metrics. It illustrates how visual data from watercolor photos may signify significant environmental alterations in aquatic habitats, presenting a promising method for efficient and economical water quality assessment.

3.3. Correlation Analysis Between Image Data and Water Quality Parameters Shimen Reservoir, Liyutan Reservoir and Hushan Reservoir

Figure 6 illustrates a detailed correlation analysis between image data ratios (G/R, B/R, G/B) and multiple water quality metrics for Shimen, Liyutan, and Hushan Reservoirs. This integrated visualization offers a comparative analysis of all three reservoirs. In Shimen Reservoir (Figure 6a), the G/R ratio exhibits significant connections with essential water quality parameters. It demonstrates a substantial positive correlation with clarity (0.83) and negative relationships with turbidity (−0.79), total phosphorus (−0.85), orthophosphate (−0.82), and dissolved phosphorus (−0.81). In Liyutan Reservoir (Figure 6b), the G/R ratio exhibits moderate to high connections with transparency (0.69) and orthophosphate (0.72). This reservoir demonstrates distinct connections between the B/R and G/B ratios with pH and temperature. The B/R ratio exhibits a negative connection with pH (−0.76) and temperature (−0.73), but the G/B ratio demonstrates positive relationships with both factors (0.73 and 0.75, respectively). At Hushan Reservoir (Figure 6c), the study indicates a robust positive association between the G/R ratio and transparency (0.80), and a negative correlation with turbidity (−0.68). This reservoir exhibits notable correlations between the B/R ratio and conductivity (0.77), alongside an inverse relationship for the G/B ratio (−0.79). These findings underscore the intricate relationship between visual data obtained from water color photographs and the physical and chemical characteristics of reservoir waters. The differing levels and patterns of association among reservoirs highlight the susceptibility of image-based monitoring methods to local environmental factors. This comparison investigation improves our comprehension of how visual criteria might signify alterations in water quality across several reservoirs, highlighting the possibilities and constraints of employing image-based techniques for water quality evaluation.

WHI Performance and Agreement with Laboratory Data

The AI-assisted WHI closely tracked seasonal water quality trends across Shimen, Liyutan, and Hushan reservoirs. For example, increasing turbidity and algal counts from February to August corresponded to declining WHI values, with most sites shifting from “Good” to “Moderate.” Cross-validation indicated strong agreement between WHI classes and laboratory-based classifications, particularly in distinguishing “Excellent/Good” from “Moderate/Poor.” Uncertainty bands (95% PIs) widened during transitional months, reflecting ecological variability. These findings demonstrate that WHI effectively condenses multiple image-derived indicators into an accessible index, enhancing interpretability for managers and communities. These prediction intervals quantify environmental variability and instrument noise, strengthening the credibility of WHI-based predictions under dynamic reservoir conditions.

4. Discussion

The results suggest that smartphone-based RGB image analysis, particularly the G/R ratio, shows promising potential as an indicator of several water-quality parameters such as turbidity, transparency, phosphorus, and algal counts. However, further validation and broader testing across different environmental conditions and devices are required before this method can be considered fully reliable for routine applications. While the smartphone-based G/R ratio demonstrates promising correlations with conventional water-quality parameters, it should be regarded as an initial diagnostic indicator rather than a definitive measurement. Various environmental factors—such as illumination variability, water-surface reflection, and differences among smartphone sensors—can affect image consistency. To address this, future work will incorporate multi-device calibration and adaptive illumination correction to enhance reliability under field conditions. The approach is intended to complement, not replace, conventional laboratory analyses, providing rapid and low-cost observations for preliminary assessment.

These findings are consistent with previous studies using remote sensing and hyperspectral imaging [17,18], which have demonstrated significant relationships between optical properties and aquatic health indicators. Compared with satellite- or laboratory-based approaches, the proposed method requires only consumer-grade devices and simple calibration, thereby reducing both costs and technical barriers.

The introduction of the AI-assisted Water Health Index (WHI) further enhances interpretability by condensing multiple predicted variables into a single score. This aligns with earlier works emphasizing integrative indices for water-quality management [18,19]. By providing near-real-time and uncertainty-aware classification, the WHI allows water managers to detect deteriorating conditions early and take proactive actions to prevent harmful algal blooms or water-quality degradation.

The limitations previously discussed, including the influence of atmospheric conditions, smartphone camera variability, and water-surface reflection, remain important considerations that may affect color-based measurements. Future research will expand testing to multiple smartphone models with different sensor architectures and image processors to ensure cross-device reproducibility. This will enable the development of calibration coefficients that standardize color response across devices, minimizing bias introduced by manufacturer-specific image processing.

Furthermore, these uncertainties can be mitigated through adaptive calibration protocols that dynamically adjust for lighting variability, viewing geometry, and atmospheric effects. Incorporating real-time illumination sensing and gray-card normalization into field procedures will help maintain consistent color capture under diverse environmental conditions.

In the broader context, the proposed smartphone-based system can be integrated into IoT-based and citizen-science monitoring networks. This integration enables real-time, community-driven observations that complement existing agency data and strengthen early-warning capabilities for turbidity or algal-bloom events [16].

Overall, while this method does not replace conventional laboratory testing, it provides an accessible and cost-effective supplementary approach for environmental monitoring. The combination of AI-assisted interpretation, spectral-ratio analysis, and low-cost imaging technology demonstrates a clear potential for democratizing water-quality assessment and expanding the spatial and temporal coverage of monitoring programs.

5. Conclusions

This study demonstrates the potential of smartphone-based imagery as a practical and affordable tool for assessing reservoir water quality. Through RGB color analysis and advanced image-processing techniques, meaningful correlations were found between water color and parameters such as turbidity, transparency, and algal concentration. Overall, the proposed method serves as a complementary and preliminary assessment tool, not a substitute for laboratory-based analysis. Its value lies in providing rapid, low-cost observations that can guide subsequent detailed testing and resource allocation in water-quality monitoring programs.

While promising, this approach should be viewed as a supplementary diagnostic technique that requires further validation across different environments and device types. Future work should focus on adaptive calibration, cross-device consistency, and integration with IoT-based sensors for real-time monitoring.

The AI-assisted Water Health Index (WHI) provides a clear framework for summarizing multiple indicators into an interpretable index, enabling wider use of smartphone imaging for participatory water-quality management.