Data from: Temporal dynamics outweigh spatial gradients in shaping water quality across a regulated Afrotropical reservoir cascade

Water quality in regulated tropical reservoirs is shaped by the interplay of seasonal hydrology, spatial habitat heterogeneity, land use, and climate variability, yet the relative importance of these drivers remains poorly quantified in African river basins. We aimed to partition the variance in water quality among temporal, spatial, land-use, and climate drivers across the Kainji–Jebba reservoir cascade in Nigeria and to identify sentinel parameters capable of consistently and timely detecting water-quality change. Monthly water-quality surveys were conducted at six sites spanning two dams and three habitat types (riverine, ecotonal, and lacustrine) from April 2024 to March 2025 (n = 72 samples). A modified National Sanitation Foundation Water Quality Index (WQI) was calculated for each sample. Variance partitioning using partial redundancy analysis (RDA), exploratory Granger causality testing, and six complementary changepoint detection methods were applied to disentangle the relative contributions of space, time, land use, and climate. Temporal variation (season and month) explained 28.3% of the variance in water quality, whereas spatial factors accounted for 1.8% and land-use effects for 0.4%. These results suggest that seasonal hydrological dynamics were the dominant structuring force within the study period, although the limited spatial extent of sampling (six sites across two reservoirs) should be considered when interpreting the comparatively weak spatial signal. After Benjamini–Hochberg false-discovery-rate correction, no climate–water-quality relationships were statistically significant across 480 stratified Granger causality tests; given the exploratory scope of the analysis and the one-year dataset, this null result is treated as hypothesis-generating rather than conclusive. WQI, electrical conductivity (EC), and relative humidity (RH) showed the highest changepoint consensus detection rates (71.4% each) and are proposed as sentinel parameters for water-quality monitoring, although the role of RH warrants further investigation before operational adoption. These findings suggest that monitoring programmes in regulated tropical reservoir cascades may benefit from prioritising targeted seasonal windows, the wet-season runoff peak (June–July) and the black flood period (January–February), rather than expanding spatial coverage, pending validation across longer time series. The integrated analytical framework, combining WQI, variance partitioning, exploratory Granger causality, and multi-method changepoint detection, offers a transferable template for other data-poor tropical river systems where monitoring resources must be allocated efficiently. Study area The study was conducted in Nigeria’s Upper Niger River Basin, centred on the Kainji–Jebba reservoir cascade (see Fig. 1 in the published article: Temporal dynamics outweigh…). Kainji Dam (10.5° N, 4.6° E; commissioned 1968) and Jebba Dam (9.1° N, 4.8° E; commissioned 1985) form a sequential hydroelectric cascade along the Niger River with a combined installed capacity of approximately 1338 MW (Chiromo et al., 2016). The region experiences a tropical savanna climate with a pronounced wet season (April–October) and dry season (November–March), receiving mean annual rainfall of 1100–1400 mm (Adegbehin et al., 2016). These seasonal transitions generate hydrological pulses that structure aquatic productivity across the cascade. At each dam, we sampled three habitat types that collectively capture the full environmental gradient imposed by impoundment: riverine zones (upstream free-flowing sections; current velocity 0.3–0.8 m s-1), ecotonal zones (transitional areas between flowing river sections and reservoir backwaters, where mixing of lotic and lentic water masses creates chemically dynamic conditions), and lacustrine zones (deep reservoir cores, > 8 m depth, with negligible current and potential for thermal stratification; Wetzel, 2001). The six sampling sites were as follows: Kainji Riverine (Garafini: 10.01555° N, 4.58584° E), Kainji Ecotonal (Awuru: 9.78447° N, 4.63231° E), Kainji Lacustrine (Yuna: 9.92093° N, 4.58925° E), Jebba Riverine (Gbajibo: 9.38147° N, 4.61988° E), Jebba Ecotonal (Juju Rock: 9.14158° N, 4.80761° E), and Jebba Lacustrine (Kokodi: 9.20615° N, 4.75372° E). Sites were georeferenced using a Garmin eTrex 30 × GPS and mapped using the sf package in R 4.5.1 (Bivand et al., 2013). Study design and temporal scope We adopted a nested sampling design, sites nested within habitat types, which were in turn nested within dams, with monthly surveys conducted from April 2024 through March 2025, yielding 12 temporal replicates at each of six sites (n=72 total samples). This hierarchical structure follows established protocols for partitioning spatial and temporal variance in lotic systems (Underwood, 1997). The design captures one complete hydrological cycle while remaining logistically feasible for a resource-constrained field programme. Each site was sampled once per month between 06:00 and 10:00 h to minimise diel variation in dissolved oxygen and temperature. All six sites were sampled within 5–7 consecutive days each month to reduce temporal confounding when assessing spatial patterns. Water quality sampling and laboratory analysis Water samples were transported to the National Institute for Freshwater Fisheries Research (NIFFFR) Water Quality Laboratory within 6 h of collection and analysed in duplicate; coefficients of variation ranged from 2.1% to 8.7%, indicating acceptable analytical precision. Surface water (0.5 m depth) was collected using a Van Dorn horizontal water sampler (Wildco, 1.2 L) at lacustrine sites accessed by motorised boat and by grab sampling at riverine and ecotonal sites. Samples were stored in acid-washed polyethene bottles at 4–6 °C during transport. Water temperature (T_w), pH, and electrical conductivity (EC) were measured in situ using a calibrated multi-parameter probe (Bluelab Combo Meter), while dissolved oxygen was measured using a portable metre (Milwaukee MW600 PRO). Water clarity was recorded as Secchi disc depth (m) and converted to a relative turbidity proxy (SD-1) for index scoring following Davies-Colley and Smith (2001); these derived values are treated as comparative proxies rather than instrument-measured nephelometric turbidity units. Nitrate-nitrogen (NO₃-N) and phosphate-phosphorus (PO₄-P) were determined by UV–Vis spectrophotometry following APHA standard colourimetric procedures (APHA, 2017). Biochemical oxygen demand (BOD₅) was determined by 5-day incubation at 20 °C. Meteorological data and land-use characterisation Monthly meteorological data: air temperature (T_a), rainfall, relative humidity, wind speed, and atmospheric pressure, corresponding to the 12-month study period (April 2024–March 2025) were obtained from stations operated by the National Institute for Freshwater Fisheries Research (NIFFR) at New Bussa (for Kainji sites) and by Mainstream Energy Solutions at Jebba (for Jebba sites), with both stations validated by the Nigerian Meteorological Agency. No multi-year climate record was available for these validated stations; accordingly, all Granger causality analyses are treated as an exploratory screening exercise within a single-year time frame rather than as a confirmatory test of long-term climatic forcing. Land-cover composition was extracted from ESA WorldCover 2021 (10-m resolution; Zanaga et al., 2022) using 5-km radius buffers around each sampling point. This radius approximates the hydrological contributing area influencing water quality at monthly sampling intervals under typical dry-season flow velocities in the Upper Niger and is consistent with buffer scales applied in previous tropical reservoir catchment studies (Carvalho et al., 2019). The percentage cover of agricultural, urban, and forest land was calculated for each buffer. Moran’s I indicated moderate spatial autocorrelation in agricultural cover (I = 0.42, p= 0.031) but not in urban or forest cover. Buffer-based land-cover metrics provide a reproducible first-order approximation of catchment conditions but do not fully capture hydrological connectivity, preferential flow pathways, or pollutant transport processes; land-use results are interpreted accordingly. Water Quality Index (WQI) calculation We calculated WQI using a modified National Sanitation Foundation method (Brown et al., 1970), aggregating eight parameters: dissolved oxygen, pH, BOD₅, T_w, NO₃-N, PO₄-P, turbidity, and electrical conductivity. Faecal coliforms were excluded because they primarily reflect public health risk rather than ecological integrity for fish. Each parameter was assigned a rescaled weight summing to 1.0 after redistributing the excluded faecal coliform weight (Table 1). Raw measurements were converted to dimensionless sub-indices (Qᵢ; 0–100) using NSF-based piecewise-linear transformation functions (Brown et al., 1970), and the composite WQI was calculated as WQI =Σ(Qᵢ×Wᵢ). A sensitivity analysis testing alternative weighting schemes (primary, proportional, and equal) confirmed that spatial and temporal WQI patterns were insensitive to moderate variations in weights (Figures S1a, S1b). WQI classes are as follows: excellent (80–100), good (65–79), medium (50–64), bad (25–49), and very bad (0–24). Differences in WQI across dams, habitat types, and seasons were assessed using Kruskal–Wallis tests with Dunn’s post-hoc comparisons and Benjamini–Hochberg false-discovery-rate (FDR) correction (Benjamini & Hochberg, 1995). Multivariate water-quality patterns were characterised using Bray–Curtis dissimilarity (Bray & Curtis, 1957) on log(x + 1)-transformed data, permutational multivariate analysis of variance (PERMANOVA; 999 permutations; Anderson, 2017), and non-metric multidimensional scaling (NMDS) ordination (stress <0.2).  Directional climate–water-quality relationships were evaluated using pairwise Granger causality tests (Granger, 1969), examining whether each of five climate drivers (atmospheric pressure, rainfall, relative humidity, air temperature, and windspeed) improved 1-month-ahead prediction of each of eight waterquality parameters (BOD, DO, EC, NO₃-N, pH, PO₄- P, turbidity, and water temperature) across all combinations of dam, habitat type, and season (2×3×2 strata=12 strata; 480 tests in total). Prior to testing, each time series was assessed for stationarity using the Augmented Dickey–Fuller (ADF) test (Dickey & Fuller, 1979); non-stationary series were first-differenced. Lag length was fixed at 1 month, with admissibility verified by minimising the Akaike Information Criterion (AIC) subject to a degrees-of-freedom constraint (maximum lag= ⌊(n- 2)/3⌋), ensuring valid F-statistics given seasonal subgroup sizes of n=5–7. Multiple testing was controlled using Benjamini–Hochberg FDR correction applied across all 480 tests. Granger causality reflects predictive temporal precedence rather than mechanistic causation. All analyses were conducted in R version 4.5.3 using the lmtest and tseries packages. Variance partitioning via partial redundancy analysis (RDA) quantified the unique adjusted R2 contributions of spatial (dam identity, habitat type), temporal (season, month), and land-use (agricultural, urban, and forest cover) predictor sets to multivariate water quality variation, using the vegan package (v. 2.6–4) in R 4.5.1 (Legendre & Legendre, 2012). Six complementary changepoint detection algorithms were applied to identify abrupt transitions in water-quality time series: pruned exact linear time (PELT), binary segmentation, variance-based detection, visual inspection of abrupt transitions, cumulative sum (CUSUM) screening, and Bayesian changepoint analysis (Zeileis et al., 2003; Killick & Eckley, 2014; Castillo-Mateo, 2022). Consensus abrupt transitions were defined as detection by two or more methods within± 1 month. Detection frequencies across parameters were used to rank candidate sentinel indicators for cost-effective monitoring. # Data from: Temporal dynamics outweigh spatial gradients in shaping water quality across a regulated Afrotropical reservoir cascade Dataset DOI: [10.5061/dryad.8pk0p2p4h](doi.org) ## Description of the data and file structure Monthly water-quality surveys were conducted at six sampling sites spanning the Kainji and Jebba reservoirs on the Niger River, Nigeria, from April 2024 to March 2025 (n = 72 samples), covering one complete hydrological cycle. Sites represented three habitat types at each dam — riverine, ecotonal, and lacustrine — capturing the environmental gradient created by impoundment. At each site and sampling month, eight physicochemical water-quality parameters were measured (dissolved oxygen, pH, biochemical oxygen demand, water temperature, nitrate-nitrogen, phosphate-phosphorus, turbidity, and electrical conductivity) and used to calculate a modified NSF Water Quality Index (WQI). Concurrent monthly meteorological data (air temperature, rainfall, relative humidity, wind speed, and atmospheric pressure) were obtained from validated stations at each reservoir. Land-cover composition (forest, agricultural, and urban cover) was extracted from ESA WorldCover 2021 within 5-km buffers around each sampling site. These data were used to partition variance in water quality among temporal, spatial, and land-use drivers; test for Granger-causal relationships between climate variables and water quality; and detect changepoints in water-quality time series using six complementary statistical methods. ### Files and variables **Water_quality_data.csv** (72 rows) — Master dataset of monthly water-quality and meteorological measurements across six sites in the Kainji–Jebba reservoir cascade, April 2024–March 2025. * `Site`: reservoir (Kainji or Jebba) * `Habitat`: habitat type (Riverine, Ecotonal, Lacustrine) * `Time_index`: sequential sampling month, 1–12 (1 = April 2024, 12 = March 2025) * `Month`: three-letter calendar month abbreviation (Apr–Mar) * `Year`: calendar year (2024 or 2025) corresponding to each Time_index * `WQI`: modified NSF Water Quality Index, unitless (0–100) * `Temp_Air`: air temperature (°C) * `RH`: relative humidity (%) * `Rainfall`: monthly rainfall (mm) * `Windspeed`: wind speed (units to confirm) * `Pressure`: atmospheric pressure (units to confirm) * `Temp_Water`: water temperature (°C) * `pH`: unitless * `DO`: dissolved oxygen (mg/L) * `BOD`: 5-day biochemical oxygen demand (mg/L) * `NO3_N`: nitrate-nitrogen (mg/L) * `PO4_P`: phosphate-phosphorus (mg/L) * `Turbidity_NTU`: turbidity, instrument-measured (nephelometric turbidity units, NTU) * `Hardness`: water hardness (units to confirm — likely mg/L as CaCO₃) * `EC`: electrical conductivity (µS/cm) **Grouped_Dataset_Kainji_Jebba.xlsx** (sheet "WQV & Meteo", 72 rows) — Same sites/months as above, formatted for time-series analysis (Granger causality and changepoint detection). Variables identical to Water_quality_data.csv except: WQI and Hardness are not included, and `Turbidity` here is a Secchi-disc-depth-derived relative turbidity proxy (SD⁻¹, dimensionless) rather than instrument-measured NTU — see Methods in the associated publication for the distinction between these two turbidity measures. **NSF_WQI_Results_Kainji_Jebba_FINAL.xlsx** (sheet "NSF_WQI_Results", 72 rows) — Intermediate WQI calculation detail, showing the sub-index score (0–100) computed for each of the eight input parameters before weighting and summation. Columns `DO_sub`, `pH_sub`, `BOD_sub`, `Temp_sub`, `Nitrate_sub`, `Turbidity_sub`, `Phosphate_sub`, `Conductivity_sub` are the individual sub-indices; `WQI` is the final weighted composite score; `WQI_Rating` is the qualitative class (e.g., Medium, Good). **Geographic_Information_Site_Map.csv** (6 rows) — Site-level metadata for the six sampling locations. * `Station`, `Ecological_Zone`, `Site_ID`, `Site_Name`: site identifiers * `Latitude`, `Longitude`: decimal degrees, WGS84 * `Date_Collected`: period during which site characterization was conducted (Aug 2022–Jul 2023; a preliminary reconnaissance survey, distinct from the water-quality sampling period of Apr 2024–Mar 2025 reported in the associated publication) * `Water_Depth_m`: approximate water depth at the site (m) * `Primary_Habitat`, `Substrate_Type`, `Additional_Notes`, `Access_Status`: descriptive site characteristics **landuse_extraction_results.csv** (6 rows) — Land-cover composition within 5-km buffers around each sampling site, extracted from ESA WorldCover 2021. * `Site`, `Habitat`, `Site_Name`: site identifiers * `forest_ha`, `agriculture_ha`, `urban_ha`: area of each land-cover class within the buffer (hectares) * `forest_pct`, `agriculture_pct`, `urban_pct`: percentage cover of each class within the buffer (%) **Missing values:** Not applicable — the dataset contains no missing values; all fields are fully populated for all 72 monthly observations and all 6 sites. ## Code/software All data files are plain CSV or Microsoft Excel (.xlsx) format and can be opened with any spreadsheet software (e.g., Microsoft Excel, LibreOffice Calc, Google Sheets) or any programming language capable of reading tabular data (e.g., Python, R). The analysis code used to process these data (variance partitioning, Granger causality testing, and changepoint detection) is openly available on GitHub: [aibrahim-art/kainji-jebba-w…](aibrahim-art/kainji-jebba-w…) All analyses were conducted in R (version 4.5.1–4.5.3), using the following packages: `tidyverse`, `dplyr`, `tidyr`, `readxl` (data handling); `vegan` (redundancy analysis, PERMANOVA, NMDS); `sf`, `rgee`, `spdep`, `ape` (spatial analysis and Google Earth Engine integration for land-cover extraction); `lmtest`, `tseries` (Granger causality testing and stationarity testing); `strucchange`, `changepoint`, `bcp`, `zoo` (changepoint detection); and `ggplot2`, `viridis`, `gridExtra` (visualization). **Workflow:** The six scripts in the GitHub repository are numbered to reflect the analysis sequence: 1. `01_land_use_extraction.R` — extracts land-cover composition from ESA WorldCover 2021 (requires site/buffer geometry files, not included in this deposit) 2. `02_WQI_computation.R` — computes the Water Quality Index from Water_quality_data.csv, producing NSF_WQI_Results_Kainji_Jebba_FINAL.xlsx 3. `03_variance_partitioning.R` — reads Water_quality_data.csv and the land-use extraction output to perform variance partitioning via redundancy analysis 4. `04a_granger_causality_full.R` and `04b_granger_heatmap.R` — read Grouped_Dataset_Kainji_Jebba.xlsx to perform stratified Granger causality testing 5. `05_changepoint_detection.R` — reads Grouped_Dataset_Kainji_Jebba.xlsx to perform multi-method changepoint detection Full methodological details are provided in the associated publication ([Temporal dynamics outweigh…](Temporal dynamics outweigh…)). ## Access information licence.Other publicly accessible locations of the data: None. This Dryad deposit is the primary and sole public location for the raw dataset, as stated in the Data availability statement of the associated publication ([Temporal dynamics outweigh…](Temporal dynamics outweigh…)). The R scripts used to analyse this dataset are separately available on GitHub: [aibrahim-art/kainji-jebba-w…](aibrahim-art/kainji-jebba-w…) (code only; no data files are hosted there). Data was derived from the following sources: * Land-cover composition was derived from ESA WorldCover 2021 (10 m resolution), a global land-cover product produced by the European Space Agency. Source: Zanaga, D., et al. (2022). ESA WorldCover 10 m 2021 v200. Zenodo. [doi.org](doi.org). ESA WorldCover is distributed under a CC BY 4.0 license. * Meteorological data (air temperature, rainfall, relative humidity, wind speed, and atmospheric pressure) were obtained from monitoring stations operated by the National Institute for Freshwater Fisheries Research (NIFFR, New Bussa) and Mainstream Energy Solutions Limited (Jebba) and validated by the Nigerian Meteorological Agency. These data were provided directly to the authors for use in this study and are not otherwise publicly licensed or distributed by the originating institutions. * All water-quality (physicochemical) measurements and site-characterisation data were collected directly by the authors during fieldwork and are original to this study.