Spatiotemporal Water Quality Assessment in Spatially Heterogeneous Horseshoe Lake, Madison County, Illinois Using Satellite Remote Sensing and Statistical Analysis (2020–2024)

Abstract

Inland lakes across the United States are increasingly impacted by nutrient pollution, sedimentation, and algal blooms, with significant ecological and economic consequences. While satellite-based monitoring has advanced our ability to assess water quality at scale, many lakes remain analytically underserved due to their spatial heterogeneity and the multivariate nature of pollution dynamics. This study presents an integrated framework for detecting spatiotemporal pollution patterns using satellite remote sensing, trend segmentation, hierarchical clustering and dimensionality reduction. Taking Horseshoe Lake (Illinois), a shallow eutrophic–turbid system, as a case study, we analyzed Sentinel-2 imagery from 2020–2024 to derive chlorophyll-a (NDCI), turbidity (NDTI), and total phosphorus (TP) across five hydrologically distinct zones. Breakpoint detection and modified Mann–Kendall tests revealed both abrupt and seasonal trend shifts, while correlation and hierarchical clustering uncovered inter-zone relationships. To identify lake-wide pollution windows, we applied Kernel PCA to generate a composite pollution index, aligned with the count of increasing trend segments. Two peak pollution periods, late 2022 and late 2023, were identified, with Regions 1 and 5 consistently showing high values across all indicators. Spatial maps linked these hotspots to urban runoff and legacy impacts. The framework captures both acute and chronic stress zones and enables targeted seasonal diagnostics. The approach demonstrates a scalable and transferable method for pollution monitoring in morphologically complex lakes and supports more targeted, region-specific water management strategies.

1. Introduction

In the U.S., there are over 123,000 lakes larger than 4 ha, plus countless smaller ones that dominate in number and collectively span approximately 9.5 million hectares [1]. These water bodies serve as ecological hotspots, providing critical services such as flood retention, recreation, and biodiversity support, while also boosting local economies and community well-being [1,2]. However, many of these lakes are increasingly threatened by nutrient pollution, sedimentation, and harmful algal blooms, often resulting from upstream land use practices and legacy contaminants [3,4]. Water pollution, particularly nutrient-driven degradation, is estimated to cost the U.S. economy over USD 4.3 billion annually in water treatment, property devaluation, and lost recreational use [5]. Most inland lakes often remain under-monitored due to their spatial complexity, seasonal variability, and the high cost of maintaining traditional in situ water quality networks [2,4]. These challenges highlight the growing need for scalable, cost-effective data collection methods and processing techniques to support proactive lake management in a changing climate [6,7].

Satellite remote sensing has significantly enhanced our ability to monitor the health of inland lakes [8,9]. Multispectral satellite missions such as Sentinel-2 and Landsat-8/9 allow for the extraction of key water quality indicators at high spatial and temporal resolutions [10,11,12]. The emergence of cloud-based platforms, such as Google Earth Engine (GEE) [13], NASA’s AppEEARS [14], and OpenET [15], has further expanded the accessibility, processing efficiency, and analytical potential of satellite-derived data. In recent years, researchers have increasingly combined these technologies to assess and visualize lake water quality across space and time. For example, Landsat imagery has been used for chlorophyll-a and organic matter retrieval [16,17,18], while Sentinel imagery has been applied to assess spatial and temporal trends in chlorophyll-a [19,20,21]. These studies have laid valuable groundwork for satellite-based lake monitoring. However, relying on a single parameter can limit our understanding of the multiple, interacting factors that contribute to water quality degradation. This is particularly true for urban inland lakes, where water conditions are shaped by a complex mix of industrial discharge, stormwater runoff, and nutrient inputs from surrounding land uses [22,23].

In response, recent research has begun to incorporate multiple indicators such as turbidity and total phosphorus to provide a more complete picture of lake health [24,25,26]. This multi-parameter approach allows for a more realistic assessment of pollution dynamics in complex lake environments [27]. However, many studies still rely on basic trend detection methods like linear regression or the Mann–Kendall test, which assume gradual changes over time [28,29,30]. As a result, they may overlook sudden pollution spikes, regime shifts, or seasonal disruptions that are critical for early warning and targeted management [31,32]. Moreover, not accounting for temporal autocorrelation can lead to misinterpretation of trend strength or significance [32]. These gaps highlight the need for more refined methods that can detect abrupt changes and capture both short-term variability and long-term patterns in water quality.

Many inland lakes, especially larger or oxbow types, are not spatially uniform. Natural landforms, engineered barriers, or variable inflow sources often divide them into distinct sub-basins. These divisions lead to localized regimes shaped by different levels of industrial discharge, stormwater runoff, agricultural input, and sedimentation. Treating such systems as homogeneous can obscure important spatial patterns and mislead intervention strategies [33,34]. In these cases, traditional time-series analyses with multiple water quality parameters may also fall short. There is a growing need to complement them with spatially aware approaches that account for intra-lake heterogeneity and detect zone-specific trends [35,36]. To better capture these internal dynamics, researchers have applied techniques such as pixel-based trend analysis [37,38], zonal averaging [39,40], moving window statistics [41,42], and region-specific comparisons [43,44]. While useful, these methods often rely on predefined spatial units and may miss shared or emergent pollution dynamics across regions [45,46,47,48]. This challenge becomes more complex when multiple water quality parameters are tracked together over long periods [6]. For effective lake management and policy planning, it is essential to move beyond univariate parameter analysis and instead synthesize multivariate time series into an integrated framework that reflects overall ecological stress. To address this, researchers have used techniques such as Self-Organizing Maps (SOMs) [49,50], Multivariate Autoregressive Models (MARs) [51,52], and index-based aggregations [53,54]. However, these methods often fall short in detecting synchronized change events across parameters and space, or in identifying the most critical time windows for intervention [55,56]. This creates a methodological gap, particularly for spatially heterogeneous lakes, that calls for more integrative and spatially dynamic approaches to pollution monitoring and decision-making.

In the current study, we address these methodological and monitoring challenges by developing an integrated framework for long-term water quality assessment in spatially heterogeneous lakes. To capture localized pollution, we divide the lake into physically distinct zones and analyze region-specific time series of chlorophyll-a (NDCI), turbidity (NDTI), and total phosphorus (TP). To overcome the limitations of basic trend methods, we apply breakpoint detection and modified Mann–Kendall tests that can capture both gradual changes and abrupt shifts. To better understand inter-zone relationships and ecological synchronization, we use correlation analysis and hierarchical clustering to detect temporal alignment and spatial similarity. Finally, to move beyond isolated trends and identify lake-wide pollution events, we employ Kernel PCA to reduce multivariate time series into a composite pollution index, which is overlaid with trend counts to pinpoint critical periods of ecological stress. Using Horseshoe Lake in Madison County, Illinois, as a case study, this framework demonstrates a scalable and transferable approach that combines satellite remote sensing, zone-wise analytics, and advanced statistical tools to support pollution detection.

2. Materials and Methods

2.1. Study Area

Horseshoe Lake is a shallow, eutrophic–turbid oxbow lake (~2 m average depth) located in the Mississippi River floodplain in Granite City, Madison County, Illinois. It lies approximately 2 miles east of the Mississippi River and about 4 miles northeast of downtown St. Louis. To capture spatial heterogeneity in turbidity and nutrient forcing, we subdivided the lake into five regions (Figure 1). Delineation used three criteria: (i) Physical separations: visible on high-resolution basemaps (levees, road embankments, and dikes that create narrow constrictions and restrict exchange); (ii) Hydrological context: proximity to major inflows/outfalls (e.g., Nameoki Ditch at the north, Elm Slough/Long Lake at the northeast, multiple agricultural drains along the east levee, occasional floodwater connections near the south/Cahokia Canal, and historical industrial outfalls on the northwest shore) [57,58]; and (iii) planform/mixing: distinct lobes with different fetch and expected residence times. The five regions are:

• Region 1 (SW lobe): Long, narrow arm isolated by a causeway; receives drainage from the south/southwest; exchanges with the main basin through a narrow opening.
• Region 2 (Central basin): Largest open-water area surrounding Walker’s Island; primary mixing zone of the lake.
• Region 3 (NW embayment): Semi-enclosed basins adjacent to the industrial complex, bounded by levees/roads; historically influenced by intake/effluent infrastructure; limited connection to Region 2.
• Region 4 (Eastern arc): Elongate eastern arm along the outer oxbow, bordered by agriculture; multiple small ditch/culvert inputs; connected to Region 2 through a constricted channel at the north.
• Region 5 (SE pocket): Shallow embayment near southern diversion structures and local drains; hydraulically restricted relative to the central basin.

These regional divisions reflect observable geomorphic and hydraulic breaks as well as known inflow points, capturing contrasting turbidity and nutrient regimes for water-quality assessment.

Figure 1. Location and Regional Zonation of Horseshoe Lake, Madison County, IL, USA.

Figure 1. Location and Regional Zonation of Horseshoe Lake, Madison County, IL, USA.

2.2. Data and Tools Used

2.2.1. Data Used

In this study, we used the Sentinel-2 Surface Reflectance Harmonized (S2_SR_HARMONIZED) dataset available on Google Earth Engine, covering the period from January 2020 to December 2024 [59]. The dataset is part of the Copernicus Program operated by the European Space Agency and, as shown in

Table 1. Sentinel-2 S2_SR_HARMONIZED Spectral Bands and Characteristics.

Band Name	Band Number	Band Description	Central Wavelength (nm)	Spatial Resolution (m)
B1	Band 1	Coastal aerosol	443	60
B2	Band 2	Blue	490	10
B3	Band 3	Green	560	10
B4	Band 4	Red	665	10
B5	Band 5	Red edge 1	705	20
B6	Band 6	Red edge 2	740	20
B7	Band 7	Red edge 3	783	20
B8	Band 8	NIR (Near-Infrared)	842	10
B8A	Band 8A	Narrow NIR	865	20
B9	Band 9	Water vapor	945	60
B11	Band 11	SWIR 1 (Short-Wave Infrared)	1610	20
B12	Band 12	SWIR 2	2190	20
QA60	-	Cloud mask bitmask	-	60

, provides multispectral imagery with spatial resolutions ranging from 10 to 60 m. With a revisit frequency of approximately five days, it enables consistent monitoring of surface water bodies. The imagery includes atmospherically corrected surface reflectance values, making it well-suited for water quality assessments over time.

In addition to satellite imagery, three geospatial datasets were incorporated to provide hydrological and landscape context for Horseshoe Lake. First, the lake boundary was defined using the National Hydrography Dataset (NHD) Waterbodies layer, which delineates surface water features across the United States [60]. Second, the American Bottom watershed boundary was obtained from the USGS Watershed Boundary Dataset (WBD), which provides hydrologically consistent watershed delineations [61]. This boundary was used to situate Horseshoe Lake within its local drainage context in the Mississippi River floodplain. Third, land use and land cover were characterized using the National Land Cover Database (NLCD 2021) at 30-m resolution, which provides nationally consistent information suitable for assessing urban and agricultural influences on water quality [62].

2.2.2. ArcGIS Pro

Esri ArcGIS Pro 3.2 is a widely used desktop GIS software (ArcGIS Pro 3.2) for spatial analysis and mapping [63]. In this study, it was used to develop and manage the spatial framework. ArcGIS Pro facilitated the creation of reference maps and the assignment of spatial attributes to satellite-derived water quality data. It also supported the export of zone-level shapefiles for use in water quality assessment within Google Earth Engine and statistical analysis in Google Colab. Overall, the platform enabled essential spatial operations and streamlined the geospatial workflow.

2.2.3. Google Earth Engine (GEE)

Google Earth Engine (GEE) is a cloud-based geospatial platform used in this study for satellite image retrieval, processing, and time-series aggregation [13]. GEE allowed efficient access to multi-temporal, multi-spectral satellite datasets. Sentinel-2 satellite data from 2020 to 2024 were fetched for the analysis. Cloud masking was performed using the QA60 bitmask approach, which excludes pixels affected by clouds and cirrus contamination to improve the accuracy of water quality parameter estimation [64,65]. Using GEE’s JavaScript interface, indices such as NDVI, NDTI, and TP were computed across the predefined lake zones. The platform significantly reduced preprocessing time and eliminated the need for local storage, streamlining the analysis.

2.2.4. Programming Interface

This study used Google Colab, a free cloud-based Python development environment, to conduct time-series analysis, principal component analysis (PCA), correlation analysis, and hierarchical clustering [66]. Python (v3.10) and widely used libraries, including Pandas (v1.5.3) [67], NumPy (v1.22.4) [68], SciPy (v1.10.1) [69], Scikit-learn (v1.2.2) [70], Matplotlib (v3.7.1) [71], and Seaborn (v0.12.2) [72], were employed to perform breakpoint detection, trend estimation, and statistical pattern recognition. This programming workflow ensured reproducible and scalable analysis and enabled seamless integration with spatial datasets from ArcGIS Pro [63] and Google Earth Engine [13].

2.3. Methodology

This study used Sentinel-2 MSI satellite imagery from 2020 to 2024 to analyze long-term spatiotemporal patterns in water quality across Horseshoe Lake. The methodology flow diagram is presented in Figure 2. Cloud-contaminated pixels were removed using the QA60 bitmask approach, ensuring clean and consistent observations. This method is widely used in remote sensing and is particularly effective for studies requiring high temporal reliability and data accuracy [64,65,73]. The resulting multitemporal image collection was used to derive key water quality indices. Three satellite-derived indicators, NDCI, NDTI, and TP, were calculated and organized into time series for each of the lake’s predefined zones [74,75,76,77,78]. These time series were used in multiple analytical steps. First, breakpoint detection was performed using the Dynamic Programming (Dynp) algorithm to identify significant changes in water quality trends [79,80]. Each segmented trend was analyzed using the Mann–Kendall test with Yue–Wang modification to classify it as increasing, decreasing, or non-significant [81,82,83,84]. These trends were then aggregated and analyzed seasonally, annually, and monthly. Second, correlation analysis was conducted on both raw time series and trend summaries to examine temporal synchronization and variability among regions. These results supported the identification of inter-index relationships and seasonal patterns. Third, hierarchical clustering was applied to group zones with similar water quality profiles, allowing for spatial classification based on pollutant behavior [85]. In parallel, PCA and trend count analysis were employed to reduce dimensionality and identify key timestamps representing high pollution periods [86]. These timestamps were used to generate representative water quality maps. Together, these integrated analyses provided a detailed understanding of water quality dynamics, pollution intensity peaks, and spatial clustering, offering valuable insights for long-term lake monitoring and management.

2.3.1. Water Quality Indices

To assess lake water quality using satellite imagery, we selected three indicators that are commonly applied in aquatic remote sensing, namely the NDCI, NDTI, and TP. These indices help quantify biological productivity, turbidity, and nutrient enrichment, respectively. We focused on chlorophyll-a (NDCI), turbidity (NDTI), and total phosphorus (TP) because together they represent the three core components of eutrophication, a leading ecological stressor in urban lakes. Unlike many other water quality parameters, these three are directly linked to both the visible symptoms (e.g., algal blooms, reduced clarity) and the underlying causes of lake degradation, and they can be reliably and cost-effectively derived from satellite imagery. This focus is particularly relevant for Horseshoe Lake, which the Illinois EPA lists on the 303(d) impaired-waters register due to documented impairments from total phosphorus, suspended solids, and recurring algal blooms [87].

• NDCI (Equation (1)) estimates chlorophyll-a concentration and detects algal blooms by combining Sentinel-2 Band 5 (red-edge, B5) and Band 4 (red, B4). Areas with high values indicate elevated phytoplankton activity [74].

$$\mathrm{N}\mathrm{D}\mathrm{C}\mathrm{I}={\displaystyle \frac{(\mathrm{B}5-\mathrm{B}4)}{(\mathrm{B}5+\mathrm{B}4)}}$$

(1)

• NDTI (Equation (2)) measures turbidity and suspended sediment levels using Sentinel-2 Band 4 (red, B4) and Band 3 (green, B3). Higher NDTI values typically correspond to poor water clarity due to sediment load [75].

$$NDTI={\displaystyle \frac{(\mathrm{B}4-\mathrm{B}3)}{(\mathrm{B}4+\mathrm{B}3)}}$$

(2)

• TP (Equation (3)) is derived using Sentinel-2 reflectance in a two-step procedure. First, Secchi depth (SD) was estimated from the red and near-infrared bands (B4, B8) with a log-linear band-ratio model shown to be effective for water clarity in turbid inland lakes [88,89]. Midwest-based studies suggest starting coefficients in the range α ≈ 2.0 to 3.0 and β ≈ −1.5 to −2.5 [16,88,89]. Based on this guidance, we adopt literature-informed priors of α = 2.5 and β = −2.2. Second, SD was mapped to the Carlson TP scale (unitless) by applying Carlson’s trophic state equations and expressing the trophic state index on the TP scale [78,90]. For brevity, we refer to this Carlson TP-scale, unitless estimate as “TP” throughout. This approach is well-suited to shallow, sediment-dominated systems such as Horseshoe Lake, where phosphorus is frequently bound to suspended particles rather than algal biomass [91]. Consistent with this context, the Illinois EPA Integrated Report Companion App lists Horseshoe Lake on the state’s 303(d) impaired-waters register in recent cycles [87].

$$TP=60-14.41 \ln (2.5-2.2 \ln ({\displaystyle \frac{\mathrm{B}4}{\mathrm{B}8}}))$$

(3)

2.3.2. Break Point and Trend Analysis

In all five regions of Horseshoe Lake, we conducted breakpoint and trend analysis using NDCI, NDTI, and TP observations from 2020 to 2024. The goal was to examine how water quality changed over time. We used the Dynp algorithm from the ruptures library to detect breakpoints [79,80]. These breakpoints mark points where significant changes occurred in the time series and are useful for identifying multiple shifts in long and noisy datasets. Each segment between the breakpoints was then analyzed using the Mann–Kendall test [81,82] with the Yue-Wang modification [83], further supported by the pyMannKendall package [84]. This version of the test corrects for autocorrelation, which helps improve the accuracy of trend detection in environmental data. Trends were categorized as increasing, decreasing, or non-significant. We calculated how long each trend type lasted in days and grouped these durations by year, season, and month. Finally, we converted the results into percentages and used bar plots to show seasonal and long-term patterns in water quality across the lake.

2.3.3. Time Series and Trend Correlations

To assess how water quality patterns vary across Horseshoe Lake, we conducted correlation analysis using both raw time-series data and long-term trends. Regional time series for key parameters (NDCI, NDTI, TP) were arranged into pivot tables, and Pearson correlation coefficients were used to measure similarity in temporal patterns between zones [92,93]. For trend analysis, we used pre-labeled directional trends (increasing, decreasing, or non-significant) from earlier time-series analyses. These trends were converted into numeric values (+1, −1, 0) and averaged by year to allow comparison across regions [94]. We then created correlation matrices to explore how consistent or different the trends were across zones. Heatmaps helped visualize these relationships, offering insight into both short-term patterns and long-term changes in water quality.

2.3.4. Hierarchical Clustering for Regional Grouping

To identify which zones of Horseshoe Lake are more or less affected by water quality issues, we applied hierarchical clustering using NDCI, NDTI, and TP time series as input features. For each zone, the median value of these parameters was calculated to represent overall conditions during the study period. The median was chosen over the mean to reduce the influence of short-term fluctuations and outliers. Before clustering, all values were standardized using z-score normalization to ensure that each parameter contributed equally to the analysis [95]. We used Ward’s linkage method, which forms clusters by minimizing the variance within each group [96]. This method was selected because it tends to produce compact, well-separated clusters, making it particularly suitable for identifying distinct patterns in environmental data across lake zones. A dendrogram was created to visualize how regions group together based on their water quality profiles. We used the Silhouette score [97] and Davies-Bouldin score [98] to evaluate the separation between clusters. The Silhouette score measures how similar each region is to its own cluster compared to others, with higher values indicating better-defined groups. The Davies–Bouldin score reflects intra-cluster compactness and inter-cluster separation, where lower values suggest more distinct and tighter clusters.

2.3.5. PCA and Trend Count Analysis for Maximum Pollution Windows

To identify the period of maximum pollution across all lake zones, we conducted a composite analysis that combined PCA with the count of increasing trend segments. Region-wise time series data were first reshaped into a consistent format containing 15 variables (three indicators across five regions). We used Kernel PCA (K-PCA) to reduce the dataset’s dimensionality and extract a single composite index (PC1) that reflects overall pollution levels [99]. K-PCA is a powerful technique for reducing dimensionality while capturing complex, nonlinear relationships in multivariate data [100]. The key parameters used in K-PCA included the kernel type, set to rbf, the number of components (n_components = 3), gamma (γ = 0.1) to control the curvature of the mapping, and coef0 = 10 to adjust the offset in the kernel function. These settings enabled effective dimensionality reduction while preserving nonlinear patterns across the 15 water quality time series. We then plotted PC1 over time and overlaid it with the number of regions simultaneously showing increasing trends. The periods where high PC1 values coincided with peak counts of increasing segments were identified as the lake’s maximum pollution periods. This integrated method allowed us to capture both the intensity and timing of water quality degradation with improved clarity.

2.3.6. Spatial Mapping for High-Pollution Windows

To assess spatial pollution patterns during the two high-pollution windows, we used ArcGIS Pro (version 3.2) [63] to calculate and map the average values of NDCI, NDTI, and TP across the five regions of Horseshoe Lake. Satellite images corresponding to the selected high-pollution periods were downloaded, and the Cell Statistics tool with the mean function was applied to compute the average of all satellite images for each pollution window. The results were then visualized as spatial distribution maps for each parameter and time window, highlighting spatial variability in pollution intensity.

3. Results

3.1. Time Series Extraction of Water Quality Indicators

Horseshoe Lake, a shallow inland water body surrounded by agricultural land and urban development, is particularly vulnerable to nutrient-rich runoff, sediment loading, and algal proliferation. These processes contribute to episodic increases in chlorophyll-a, turbidity, and phosphorus, key indicators of eutrophication and declining water quality. To evaluate these dynamics, the current study utilized NDCI, NDTI, and TP indices. A five-year stack of Sentinel-2 MSI imagery (2020–2024) was processed in Google Earth Engine. Cloud-contaminated pixels were masked using the QA60 bitmask, and image data were spatially extracted for each of the lake’s five zones. For each zone, we computed mean index values per image, generating time series with a nominal 5-day revisit interval. Figure 3 presents the time series of NDCI, NDTI, and TP across all zones. These subplots reveal distinct seasonal patterns and inter-annual variability in water quality. Notably, certain regions exhibit consistently higher turbidity (region 5) or chlorophyll-a (region1) concentrations, reflecting localized sources of disturbance or stagnation. The QA60 masking was validated through visual inspection and temporal stability checks during known cloudy periods, ensuring reliable time series inputs for subsequent analyses. To assess the distribution and variability of water quality conditions, we generated box-and-whisker plots (Figure 4) for NDCI, NDTI, and TP across the five lake zones. Region 1 had the highest NDCI median (0.083) and upper quartile (Q3) value (0.2065), indicating elevated chlorophyll-a concentrations, while Region 5 recorded the lowest median (0.0305). NDTI medians followed a similar trend, with Region 1 at −0.0510 and Region 5 close to zero (−0.0025), suggesting clearer water in the lake’s peripheral zones. TP values peaked in Region 5, with a median of 0.1875 and a maximum of 0.412, while Region 2 had the lowest median (0.0460). Overall, Regions 1 and 5 consistently showed higher pollution levels across all indicators, marking them as potential hotspots for eutrophication and turbidity.

3.2. Trends in Water Quality Parameters

Figure 5, Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10 present a comprehensive assessment of water quality trends in Horseshoe Lake from 2020 to 2024 using NDCI, NDTI, and TP, respectively. For each parameter, two figures are provided. The first shows region-specific time series with detected breakpoints and trend segments classified as increasing, decreasing, or non-significant. The second summarizes the distribution of these trends at annual, seasonal, and monthly levels. Together, these visualizations illustrate both spatial (region-wise) and temporal changes in water quality.

Figure 3. Time series of NDCI (green), NDTI (orange), and TP (blue) across five regions of Horseshoe Lake from 2020 to 2024.

Figure 3. Time series of NDCI (green), NDTI (orange), and TP (blue) across five regions of Horseshoe Lake from 2020 to 2024.

Figure 4. Box-and-whisker plots showing the distribution of NDCI, NDTI, and TP values across the five regions of Horseshoe Lake. Colors indicate different regions: Region 1 = teal, Region 2 = orange, Region 3 = blue, Region 4 = pink, and Region 5 = green.

Figure 4. Box-and-whisker plots showing the distribution of NDCI, NDTI, and TP values across the five regions of Horseshoe Lake. Colors indicate different regions: Region 1 = teal, Region 2 = orange, Region 3 = blue, Region 4 = pink, and Region 5 = green.

• Trends in NDCI: Figure 5 shows NDCI trend segments across five lake regions. Region 1 experienced a fairly balanced sequence of increasing and decreasing trends, with a few non-significant periods. It shows recurring fluctuations, especially between 2021 and 2023. Region 2 started with short-term declines, followed by frequent alternating increases and decreases. Region 3 showed higher variability, with short trend segments and more frequent declines during 2021–2023. Region 4 had longer periods of consistent decline, especially from mid-2020 to late 2022, with limited signs of recovery. In contrast, Region 5 experienced some of the longest periods of both increase and decrease. It showed extended rises in NDCI during 2022 and early 2024, followed by a decline through the end of the study period. Figure 6 summarizes trend distributions across annual, seasonal, and monthly scales for NDCI. Annually, decreasing trends dominated in 2020 and 2023, while 2021 and 2022 showed more frequent increases. Seasonally, fall had the strongest NDCI declines, with 86% of periods showing decreasing trends. Spring and summer displayed a mix of increases and decreases. Winter recorded the highest share of increasing trends at 58%. Monthly patterns followed these trends, with February and July showing peaks in increases, while September and October were entirely marked by declines.

Figure 5. Time series of NDCI with detected breakpoints and classified trend segments (2020–2024) for the five Horseshoe Lake regions. Trend colors indicate direction: red = increasing, green = decreasing, and gray = non-significant.

Figure 5. Time series of NDCI with detected breakpoints and classified trend segments (2020–2024) for the five Horseshoe Lake regions. Trend colors indicate direction: red = increasing, green = decreasing, and gray = non-significant.

Figure 6. Distribution of NDCI trend directions in Horseshoe Lake from 2020 to 2024. Bar plots summarize the relative proportion of increasing (red), decreasing (green), and non-significant (gray) trends based on duration-weighted segments, shown across annual, seasons, and months.

Figure 6. Distribution of NDCI trend directions in Horseshoe Lake from 2020 to 2024. Bar plots summarize the relative proportion of increasing (red), decreasing (green), and non-significant (gray) trends based on duration-weighted segments, shown across annual, seasons, and months.

• Trends in NDTI: Figure 7 shows NDTI trends from 2020 to 2024 across the five lake regions. Region 1 had a mix of trends, with several short periods of increase and a few longer decreasing segments, showing alternating turbidity behavior. Region 2 showed mostly increasing trends early on, but more decreasing periods appeared between 2021 and 2023. A few increases returned in 2024. Region 3 was the most dynamic, with many short segments and a balance of increases and decreases. However, there was a cluster of persistent increases from late 2022 through 2024. Region 4 was dominated by long periods of decreasing turbidity from 2021 to 2023, followed by several shorter increases, suggesting recovery followed by new disturbances. Region 5 had the most consistent increases, especially in 2020, late 2022, and throughout 2024. Figure 8 shows NDTI trends by year, season, and month. In 2020 and 2023, increasing trends were most common, reaching up to 59%. In 2021 and 2024, decreasing trends were more frequent, reaching 55% to 58%. Spring had the highest share of decreasing trends at 74%. Fall showed the most increasing trends at 77%. Summer had a mix of both. Winter showed nearly equal shares of increases and decreases. At the monthly level, April and May had the strongest decreases, with up to 88%. August, October, and November showed the highest increases.
• Trends in TP: Figure 9 shows TP trends in Horseshoe Lake from 2020 to 2024. Region 1 had mostly increasing trends throughout the period, with short declines in late 2020 and mid-2021. Region 2 showed a mix of patterns, with early increases, mid-period declines, and more increases in 2024. Region 3 started with mostly increasing and non-significant trends, but showed consistent declines in mid to late 2022 and again in 2024. Region 4 had an early increasing phase, followed by a long declining trend from mid-2021 to late 2023, then returned to short increases and stable periods. Region 5 showed the most prolonged and consistent increases, especially from early 2020 and again in late 2023 to the end of 2024, with only a few brief declining periods. Figure 10 summarizes annual, seasonal, and monthly TP trends. In 2020, increasing trends were highest at 59%. In 2021 and 2022, decreasing trends were more common, peaking at 66% in 2022. Increases returned in 2023 and 2024. Summer had the highest share of decreasing trends at 64%. Fall showed the most increasing trends at 72%. Spring and winter had more balanced patterns. Monthly trends followed this pattern. June and September had the strongest decreases. October and November showed the highest increases, close to 100%. February and August had more non-significant trends.

Figure 7. Time series of NDTI for the five Horseshoe Lake regions with Breakpoints and Trend Lines from 2020 to 2024. Trend colors indicate direction: red = increasing, green = decreasing, and gray = non-significant.

Figure 7. Time series of NDTI for the five Horseshoe Lake regions with Breakpoints and Trend Lines from 2020 to 2024. Trend colors indicate direction: red = increasing, green = decreasing, and gray = non-significant.

Figure 8. Distribution of NDTI trend directions in Horseshoe Lake from 2020 to 2024, based on duration-weighted segments across annual, seasons, and months.

Figure 8. Distribution of NDTI trend directions in Horseshoe Lake from 2020 to 2024, based on duration-weighted segments across annual, seasons, and months.

Overall, the trends in NDCI, NDTI, and TP show both short-term variability and longer-term shifts in water quality across Horseshoe Lake. These patterns reflect seasonal cycles, spatial differences, and the need for continued monitoring to guide local management and restoration strategies.

Figure 9. Time series of TP for the five Horseshoe Lake regions with Breakpoints and Trend Lines from 2020 to 2024. Trend colors indicate direction: red = increasing, green = decreasing, and gray = non-significant.

Figure 9. Time series of TP for the five Horseshoe Lake regions with Breakpoints and Trend Lines from 2020 to 2024. Trend colors indicate direction: red = increasing, green = decreasing, and gray = non-significant.

Figure 10. Distribution of TP trend directions in Horseshoe Lake from 2020 to 2024, based on duration-weighted segments across annual, seasons, and months.

Figure 10. Distribution of TP trend directions in Horseshoe Lake from 2020 to 2024, based on duration-weighted segments across annual, seasons, and months.

3.3. Time Series and Trend Correlations

Figure 11 summarizes correlation matrices for both raw time series and trend-segmented values of NDCI, NDTI, and TP across the five lake regions. For the raw time series, NDCI and TP show consistently high correlations across most region pairs. In contrast, NDTI displays more varied and weaker associations. TP raw values exhibit strong alignment among all regions, with most correlation coefficients above 0.9. This reflects the lake-wide influence of nutrient-rich inflows and sediment resuspension that broadly affect phosphorus concentrations. NDCI raw values also show moderate to strong agreement, particularly between Region 2 and Regions 3 and 5. This suggests shared biological responses in adjacent zones influenced by common hydrological inputs. NDTI raw correlations drop substantially for Region 5 when compared with other zones. This is likely due to persistent turbidity stress from multiple agricultural culverts and shallow bathymetry that create localized sediment resuspension dynamics. For trend-based values, the correlation structure is more differentiated. NDCI trend correlations cluster more strongly among adjacent regions, such as Regions 2 and 3 or Regions 3 and 4. This shows that phytoplankton activity is synchronized in hydrologically connected zones. Isolated areas respond more independently. NDTI trend relationships are concentrated between Regions 2, 3, and 4. Correlations are near zero elsewhere. This pattern is consistent with shared stormwater inputs and sediment dynamics in the eastern and central portions of the lake. TP trend correlations remain moderate to high among Regions 1, 2, and 3. These areas receive steady nutrient inputs from urban and industrial inflows. Region 4 shows weaker alignment with others, reflecting its relative isolation and different inflow sources. Overall, these patterns indicate that while phosphorus and chlorophyll-a are influenced by lake-wide drivers, turbidity is more localized and spatially heterogeneous, tied closely to site-specific inflows and sediment conditions.

3.4. Hierarchical Clustering for Regional Grouping

As shown in Figure 12, hierarchical clustering grouped the five regions of Horseshoe Lake into three clusters based on their median values of NDCI, NDTI, and TP. The dendrogram illustrates similarity among regions, where lower linkage distances indicate stronger resemblance in water quality profiles. Regions 1 and 2 clustered together at a distance of 0.8, reflecting comparable conditions with moderate chlorophyll-a, low turbidity, and low phosphorus. Regions 4 and 5 merged at 1.3, both characterized by low chlorophyll-a but elevated turbidity and phosphorus. Region 3 joined the 1–2 cluster at 2.6, indicating intermediate conditions that partly overlap with the urban-influenced zones. The final division occurred at 4.2, separating the (1–2–3) group from the (4–5) group and highlighting a distinct nutrient–turbidity regime in the southeastern portion of the lake. Cluster validity indices (Silhouette = 0.26, Davies-Bouldin = 0.42) indicate moderate separation, suggesting that while clusters are distinguishable, inter-regional overlap persists due to shared but unevenly distributed pollution drivers.

3.5. PCA & Trend Count Analysis for Maximum Pollution Windows

To identify representative windows of maximum pollution across Horseshoe Lake, we first reduced the three core indicators, NDCI, NDTI, and TP, into a single composite stress index using Kernel PCA (K-PCA) (Figure 13). The first principal component (PC1) explained ~80% of the total variance, making it a robust summary of overall water-quality stress. We then screened the time series for periods when both (i) the composite stress index was elevated and (ii) the number of lake regions showing increasing trends was high. This dual criterion ensured that selected events were not only intense but also spatially coherent across the lake, thereby suitable for examining spatial heterogeneity. Two distinct windows met these conditions:

• 31 December 2022–15 January 2023: PC1 values ranged from 1.33 to 2.26, with a peak of 2.26 on 3 January 2023. Values were ≥2.0 from 31 December 2022 through 8 January 2023. This indicates strong pollution across chlorophyll-a, turbidity, and phosphorus. The increasing-segments count reached 11 (highest in the full series), suggesting a fast, steady rise in pollution indicators.
• 24 November 2023–10 January 2024: PC1 values rose from −0.51 (24 November 2023) to a peak of 2.67 (10 January 2024), remaining >2.0 through early January (≈2.00–2.67). The increasing-trend count stayed between 9 and 10, indicating a longer-lasting pollution event with steady upward changes in water-quality indicators.

This analysis identified two distinct pollution windows for further study. The first (31 December 2022–15 January 2023) reflects a short but sharp spike in pollution affecting nearly all lake regions simultaneously, while the second (24 November 2023–10 January 2024) represents a longer, more moderate yet persistent event. Together, these contrasting episodes provide a basis for examining how spatial variation in water quality develops under different pollution dynamics in Horseshoe Lake.

Figure 13. Composite Pollution Metric (PC1) and Increasing Trend Segments Over Time (2020–2024), Including Explained Variance Ratio from K-PCA Analysis.

Figure 13. Composite Pollution Metric (PC1) and Increasing Trend Segments Over Time (2020–2024), Including Explained Variance Ratio from K-PCA Analysis.

3.6. Regional Water Quality Patterns During Pollution Peaks

Spatial distributions of NDCI, NDTI, and TP during the two identified pollution windows provide insights into how different regions of Horseshoe Lake responded to peak water quality stress. These patterns are illustrated in Figure 14a (for the first window: 31 December 2022–15 January 2023) and Figure 14b (for the second window: 24 November 2023–10 January 2024).

• 31 December 2022–15 January 2023: The NDCI map shows high chlorophyll-a levels in Region 1 and Region 5 (green areas), indicating strong algal activity. Region 2 has moderate values, mostly in its southern part. Region 3 records the lowest NDCI (purple), suggesting clearer water. Region 4 shows a mix of low and moderate values. These patterns suggest that biological stress was highest in the southern and southeastern zones, aligning with the PC1 pollution peak. The NDTI map highlights elevated turbidity in Region 5, likely from sediment or surface runoff. Region 4 has small patches of moderate turbidity. Regions 1, 2, and 3 mostly show low values (purple), reflecting clearer conditions. This suggests turbidity stress was concentrated in Region 5. TP values were also highest in Region 5 and parts of Region 4. These areas likely received nutrients from nearby agriculture or disturbed sediments. In contrast, Regions 1, 2, and 3 show low phosphorus levels. Together, these findings show that nutrient and turbidity-related pollution was localized in the southeastern part of the lake.
• 24 November 2023–10 January 2024: During this period, high NDCI values appear in Region 1 and parts of Region 5, indicating strong algal growth. Region 3 has the lowest chlorophyll-a levels, while Regions 2 and 4 show moderate values with a few high-value patches. The spatial spread points to increased biological stress in the southern and southeastern lake zones. Turbidity was again highest in Region 5, shown by green areas on the NDTI map. Region 4 has moderate turbidity, while the rest of the lake (Regions 1, 2, and 3) shows lower values. This indicates that physical disturbance was concentrated in Region 5. The TP map shows a similar pattern. Regions 4 and 5 had the highest phosphorus levels, suggesting nutrient inputs from runoff or sediments. The other regions remained low in TP. The overlap of high NDCI, NDTI, and TP confirms a strong, localized pollution hotspot in the southern zones during this window.

3.7. Validation and Uncertainty

Although direct field measurements of chlorophyll-a, turbidity, and phosphorus were not collected during our study, we validated our satellite-derived indices against independent evidence from the literature and local observations. To address this, we adopted a multi-source validation framework. First, the NDCI, NDTI, and TP are well established in aquatic remote sensing studies and have been calibrated in turbid, eutrophic lake environments similar to Horseshoe Lake [16,74,89]. Second, recent field-based investigations in Horseshoe Lake provide convergent support for our findings. Ojewole (2023) reported fall peaks of fecal-indicator bacteria (E. coli and Enterococci) in October–November 2022 [101]. These peaks correlated positively with elevated concentrations of EC, TDS, salinity, alkalinity, and hardness. This confluence of signals, high bacteria and high water chemistry indicators, is consistent with the high-turbidity conditions captured by our NDTI and the nutrient-enriched context implied by our TP observations over the same seasonal window. Dimpor et al. (2025) monitored Horseshoe Lake from December 2022 to November 2023 and reported consistently high detection of fecal indicator bacteria, with enterococci concentrations often surpassing EPA recreational thresholds during summer months [102]. Their analysis highlighted water temperature, precipitation, salinity, resistivity, dissolved oxygen, and hardness as significant predictors of bacterial variability. These findings correspond with our NDCI observations of elevated biological productivity in summer–fall, our NDTI detection of turbidity increases linked to ion-rich conditions, and our TP estimates showing nutrient enrichment during runoff-driven periods. Third, the Illinois EPA 303(d) impairment listing identifies turbidity and nutrient enrichment as major impairments in Horseshoe Lake, consistent with both our results and these independent field studies [87]. Finally, qualitative observations provided by the lake superintendent confirmed the persistence of bloom-prone, low-clarity zones in areas highlighted by our satellite indices. Together, these lines of evidence confirm that our satellite-derived indices reliably capture the seasonal and spatial dynamics of turbidity and nutrient stress in Horseshoe Lake, despite the absence of direct in situ measurements.

4. Discussion

This five-year study of Horseshoe Lake examines how hydrological, ecological, and human drivers shape water quality in a shallow, eutrophic–turbid floodplain system. In interpreting the results, we draw on three explanatory hypotheses to understand the observed patterns: (i) urban inflows drive persistent chlorophyll-a enrichment in regions directly exposed to stormwater culverts, (ii) agricultural discharges sustain elevated turbidity and phosphorus levels in zones surrounded by croplands, and (iii) seasonal latent stress emerges in fall and winter, when turbidity and nutrient levels remain high even in the absence of strong algal activity. Using Sentinel-2 satellite imagery in combination with breakpoint detection, trend tests, and spatial clustering, we explored these hypotheses by analyzing both temporal change and spatial heterogeneity in chlorophyll-a, turbidity, and total phosphorus indicators. The results provide new insights into how land use-land cover context and seasonal drivers interact to influence water quality dynamics. These insights offer an evidence base for long-term monitoring, predictive modeling, and adaptive lake management.

• Regional Drivers of Pollution and Spatial Heterogeneity: Horseshoe Lake is situated within the American Bottom watershed, a floodplain of the Mississippi River where land use land cover is dominated by urban development and agriculture (Figure 15) [60,61]. Our results validate the urban inflows hypothesis and the agricultural discharge hypothesis, showing that different land use-land cover contexts leave distinct ecological signatures in the lake. The consistently high chlorophyll-a levels in the north are best understood as a consequence of stormwater culverts draining Granite City into Region 1, aligning with elevated NDCI values and repeated chlorophyll-a spikes. On the eastern margin, croplands surround Region 5 and deliver multiple agricultural discharges, producing persistent hotspots of turbidity and phosphorus enrichment. The combination of nutrient-rich inflows, shallow bathymetry, and sediment resuspension reinforces a chronic stress regime evident across multiple indicators and time windows. The industrial legacy on the western shore has diminished since effluent discharges ceased [57,58]. As a result, the lake’s dominant external drivers are now stormwater, agricultural runoff, and seasonal snowmelt. Inflows from Elm Slough, Long Lake, and the Cahokia Drainage Canal further reinforce watershed-lake connectivity, producing synchrony among central and eastern regions. These patterns demonstrate that land use, hydrology, and connectivity jointly structure water quality outcomes. Although these findings are specific to Horseshoe Lake, they can be generalized to other shallow floodplain lakes across the American Midwest. In such systems, urban, agricultural, and historical industrial inputs combine to produce highly variable water quality patterns across different areas of the lake [16,88,89]. More broadly, they illustrate a global principle for eutrophic-turbid lakes [91]. Region-specific management is often more effective in these systems than uniform lake-wide interventions.
• Temporal Disruption and Seasonality in Trends: Breakpoint detection confirmed that Horseshoe Lake’s water quality dynamics are nonlinear and event-driven, validating the seasonal latent stress hypothesis. Abrupt shifts were frequently triggered by storm events, flood diversions, or seasonal nutrient pulses [103,104]. Seasonal summaries showed that fall and winter carried elevated risks, with high turbidity and phosphorus persisting even when algal activity was low [105]. Such “latent stress” periods illustrate the hidden dimensions of eutrophication that can be missed by summer-centric monitoring [103,106]. This finding has broad implications for Midwestern lakes, where freeze–thaw cycles, storm-driven runoff, and nutrient flushes occur outside of the summer growing season [107,108]. Globally, it resonates with floodplain and monsoon-influenced lakes where episodic events dominate water quality trajectories. The general lesson is that year-round, event-sensitive monitoring frameworks are essential to accurately capture eutrophication stress in turbid, nutrient-rich systems.
• Inter-Zonal Synchronization and Spatial Clustering: Spatial clustering analysis further supported the urban inflows and agricultural discharge hypotheses, revealing that Horseshoe Lake is not a uniform system but a mosaic of zones shaped by distinct external drivers. Regions 1 and 2 clustered together under the influence of urban runoff, while Regions 4 and 5 consistently grouped under agricultural and turbidity pressures. Region 3 stood apart as a transitional buffer, reflecting mixed inputs but reduced stress. These patterns emphasize that sub-regional variation is not noise but an organizing feature of eutrophic–turbid lakes. Similar clustering has been documented in other Midwestern floodplain lakes [109]. These patterns show that land use-land cover and hydrological setting create consistent and predictable water quality signatures [104]. Beyond the Midwest, this principle holds in shallow lakes worldwide, where spatial heterogeneity means that lake-wide averages can obscure critical dynamics [110,111]. For both science and management, the implication is clear [109,112]. Tailored, zone-specific strategies are more effective than one-size-fits-all approaches.
• Framework Utility and Broader Application: Beyond site-specific findings, our results also demonstrate the value of an integrated methodological framework for monitoring eutrophic and turbid lakes such as Horseshoe Lake. The combination of Sentinel-2 imagery with key water quality indicators, NDCI, NDTI, and TP, proved effective in capturing the dominant stressors in this shallow floodplain system [16,78,90]. Breakpoint detection highlighted abrupt disruptions in these indicators, contrasting with earlier studies that largely relied on seasonal or annual averages to describe trophic dynamics [113,114,115]. More recent work has similarly shown the value of event-focused detection in lake time series, supporting our approach [116,117]. In parallel, non-parametric trend and correlation analyses provided robust assessments of temporal variability, offering a more sensitive alternative to the linear regression approaches commonly applied in past limnological studies [81,82,83]. Similar non-parametric tools have gained traction in recent aquatic monitoring studies, reinforcing their suitability for complex lake dynamics [118]. The integrative use of PCA with trend counts further allowed us to identify periods of maximum pollution intensity by combining signals across indicators. This extends beyond single-variable thresholds typically applied in trophic state monitoring [119] and aligns with more recent applications of multivariate methods in inland water quality assessments [120,121]. Finally, hierarchical clustering exposed distinct spatial groupings within the lake, underscoring heterogeneity that would have been masked if lake-wide averages were the sole basis of assessment [122,123]. Taken together, these results show that the framework is capable of detecting both short-term pollution spikes and longer-term structural patterns within a single pipeline. This provides a finer-grained, event-sensitive view of water quality than earlier approaches, which tended to emphasize either individual indicators or coarse temporal summaries [114,115]. In sum, this integrated pipeline offers a transferable, decision-support tool for eutrophic–turbid lakes, pinpointing when and where to focus monitoring and guiding cost-effective, region-specific management actions. Its scalability is strongest in lakes with frequent satellite coverage, shallow depths, and strong land-water interactions where episodic events drive water quality dynamics. By contrast, its application is more limited in optically complex systems dominated by submerged vegetation, in high-latitude regions with persistent cloud and ice cover, or in deep stratified lakes where surface reflectance poorly represents nutrient dynamics.

Taken together, these results support our three interpretive hypotheses and demonstrate that hydrological, ecological, and human drivers interact in complex but explainable ways in eutrophic-turbid lakes. The patterns observed in Horseshoe Lake reflect processes common to many shallow floodplain systems in the American Midwest, where urban inflows, agricultural runoff, and legacy sediments jointly shape water quality. At the same time, the event-driven disruptions, seasonal latent stress, and spatial heterogeneity documented here resonate with shallow lakes worldwide, from monsoon-influenced floodplains to agriculturally dominated watersheds. By framing our interpretations through these hypotheses within a flexible, satellite-based framework, this study provides not only site-specific insights but also a transferable approach for monitoring, modeling, and managing eutrophic–turbid lakes under changing climatic and land use pressures.

5. Conclusions

This five-year spatiotemporal assessment of Horseshoe Lake demonstrates the value of integrating satellite remote sensing with advanced statistical techniques to monitor inland lake water quality in morphologically complex systems. The analysis revealed persistent pollution hotspots in Regions 1 and 5, reflecting their unique exposure to external stressors. Region 1 is directly influenced by stormwater culverts draining Granite City, delivering concentrated urban runoff that fuels recurring algal activity. Region 5, by contrast, receives multiple agricultural discharges along its eastern margin and is further shaped by shallow bathymetry, which promotes sediment resuspension and sustained turbidity. Other regions did not exhibit comparable stress levels because they are less hydrologically connected to these intensive inflow sources or act as transitional zones with relatively lower nutrient and sediment loading. Temporal analysis captured both gradual and abrupt shifts in key indicators. Breakpoint detection and the modified Mann–Kendall test successfully revealed pollution spikes and seasonal stress events that are often missed by conventional linear trend methods. Notably, increasing trends in turbidity and phosphorus were most common during fall and winter, suggesting that summer-centric monitoring efforts may overlook ecologically significant periods of risk. Through spatial correlation and hierarchical clustering, the study identified distinct zone-level pollution regimes, reinforcing the need for targeted, region-specific management strategies rather than a one-size-fits-all approach. The integration of Kernel PCA with trend segmentation further enabled the detection of two critical pollution periods, late 2022 and late 2023, offering a lake-wide perspective on ecological stress and demonstrating the value of multivariate synthesis in capturing synchronized water quality responses. This study contributes to the field of satellite remote sensing by combining zone-specific analysis, multivariate integration, and trend segmentation into a unified, scalable framework. It advances beyond traditional multi-parameter monitoring by offering a more nuanced, time- and space-sensitive method for assessing ecological stress in spatially heterogeneous inland lakes. Importantly, the indices employed here were validated through convergent evidence from literature, regulatory assessments, and local observations, even in the absence of co-located in situ matchups. While this provides confidence in the reliability of our results, future incorporation of direct field-based calibration and validation would further strengthen the quantitative inferences drawn. Looking ahead, future work should also link water quality dynamics more explicitly to hydrological and meteorological drivers such as precipitation events, land use changes, and inflow variability. Operationalizing this framework for other inland lakes, particularly in data-scarce or resource-limited settings, could further support proactive monitoring and early-warning systems through automation and cloud-based platforms. Together, these insights establish a robust and transferable model for advancing inland water quality assessment and management using remote sensing technologies.