Harmful Algal Blooms (HABs) pose a significant threat to freshwater ecosystems and public health globally, necessitating reliable early warning systems for effective water resource management. This study presents an end-to-end AI framework designed to predict 4-level ordinal algal alerts in South Korea by systematically integrating heterogeneous spatio-temporal environmental datasets, including GIS-based spatial features, water quality, meteorological, and hydrological data. Our methodological approach involves: (1) extracting spatial features via GIS; (2) optimizing time-lags and interpolating time-series data based on Spearman correlation; and (3) performing ordinal classification using a LightGBM (LGBM) model. To address the ordinal nature of algal alerts, the model was optimized using Optuna with the Quadratic Weighted Kappa (QWK) metric. A rigorous Dual Cross-Validation (CV) framework was employed to assess generalization capabilities: Year-over-Year (YoY) CV with an Embargo technique was used to evaluate temporal performance while preventing data leakage, and Leave-One-Station-Out (LOSO) CV was applied to validate spatial generalization for unobserved locations. Additionally, Isotonic Regression was implemented for probability calibration to enhance the reliability of the predicted outputs. By effectively controlling spatio-temporal information leakage, this study demonstrates superior predictive performance across unobserved timeframes and locations, providing a robust decision-support tool for practical water quality management.

How to cite: Lee, S., Lee, H., Baek, S., Necesito, I. V., and Kim, S.: Robust Ordinal Algal Alert Prediction Framework Integrating Heterogeneous Spatio-temporal Data and Dual Cross-Validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16503, https://doi.org/10.5194/egusphere-egu26-16503, 2026.

Accurate prediction of nitrate‑nitrogen (NO₃‑N) and total phosphorus (TP) loads is critical for managing water quality in agricultural watersheds, where excess nutrient runoff can contribute to downstream eutrophication. This study applies Random Forest (RF) regression and conditional inference RF to predict monthly NO₃‑N and TP loads at eight monitoring gages within Conservation Reserve Enhancement Program (CREP) watersheds in the Illinois River and Kaskaskia River basins. The machine learning (ML) models were trained using hydroclimatic, land‑use, and nutrient datasets from 2000–2022 and validated with 2023 observations.
The predictor variables included discharge, precipitation, temperature, land use, total Kjeldahl nitrogen (TKN), suspended sediment, and septic system density in the watersheds. Multiple modeling strategies were evaluated, including full‑feature, reduced‑feature (i.e., derived through importance thresholds or removal of collinear nutrient variables), and hyperparameter‑tuned configurations. Model performance was assessed using Nash–Sutcliffe Efficiency (NSE), R², and RMSE, and interpretability was evaluated through feature‑importance metrics and SHAP analyses.
Monthly Random Forest models effectively captured seasonal nutrient dynamics. Discharge consistently emerged as the dominant predictor of NO₃‑N loads, while interactions among variables, particularly TKN and suspended sediment, played major roles in predicting TP. Land use and septic system density exhibited limited predictive influence. Model performance was strong across configurations, with training NSE values exceeding 0.95 and validation NSE frequently above 0.9. However, reduced skill during summer and fall suggested the influence of unrepresented processes such as evapotranspiration. The most stable performance across sites and seasons was achieved with hyperparameter‑tuned, full‑feature models. SHAP analyses revealed clear linkages between hydrologic and biogeochemical processes, while Spearman correlation heatmaps highlighted strong covariation among nutrient loads and moderate coupling with climatic variables.
These results demonstrate the value of machine‑learning approaches such as RF as complementary alternatives to process‑based models like SWAT, offering robust tools for informing nutrient‑reduction strategies and supporting policy decisions in impaired agricultural watersheds.

How to cite: Getahun, E. and Kharosekar, R.: Data‑Driven Modeling of Nutrient Dynamics: Random Forest Predictions of Nitrate and Total Phosphorus Loads in Illinois, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16874, https://doi.org/10.5194/egusphere-egu26-16874, 2026.

Traditional hydrological models have been widely applied to flood simulation across the globe, yet the accurate simulation of peak discharge remains a long-standing shortcoming of these models. Taking the upper reaches of the Dongjiang River and Beijiang River in South China as the study area, this study employed the Variable Infiltration Capacity (VIC) model to capture the peak discharge during flood events. The simulation period was divided into a calibration period (2011–2014) and a validation period 1 (2008–2010) before vegetation changes, as well as a validation period 2 (2015–2020) after vegetation changes. The results demonstrated that the VIC model exhibited good applicability in both the upper Dongjiang River and upper Beijiang River basins. For the upper Beijiang River basin, the Nash-Sutcliffe Efficiency (NSE) and Kling-Gupta Efficiency (KGE) values were both above 0.6 during the calibration period, while these values were close to 0.6 in both validation periods 1 and 2. However, the model consistently underestimated the peak discharge in all periods. To address this limitation, a machine learning approach was introduced by coupling the VIC model with the Bidirectional Long Short-Term Memory (Bi-LSTM) network. Specifically, the soil moisture content, grid-scale runoff simulated by the VIC model, and precipitation data were used as training inputs for the Bi-LSTM model. Meanwhile, the standalone VIC model and pure Bi-LSTM model were set as control groups for comparison. The results indicated that the coupled VIC-Bi-LSTM model outperformed the control groups in capturing both the runoff process and peak discharge in the two basins. During the calibration period, the NSE values of the coupled model reached 0.9, and remained above 0.7 in both validation periods. In addition, scenarios before and after vegetation changes were designed to analyze the performance of the VIC model in simulating runoff under varying underlying surface conditions. The results revealed that the VIC model could effectively capture the impacts of vegetation changes on runoff, with the NSE value in validation period 2 (post-vegetation change scenario) being close to that in validation period 1. Moreover, the coupling with Bi-LSTM enabled more precise simulation of runoff in the upper Dongjiang and Beijiang Rivers under the scenario of altered vegetation cover.

How to cite: Han, G., He, Z., and Sun, H.: Coupling Machine Learning with Physical Models to Improve Peak Flood Simulation under Vegetation and Rainstorm Variability, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7108, https://doi.org/10.5194/egusphere-egu26-7108, 2026.

Transit time to baseflow refers to the amount of time between when a water parcel enters a catchment as precipitation and when it exits the system via discharge. It is a key concept that links climate variability, hydrological transport, and biogeochemical processes, with broad implications for both surface water and groundwater quality, resource sustainability, and vulnerability to climate change impacts. Transit time distributions can be inferred from spatially resolved time-series measurements of environmental tracer concentrations, but such observations are typically available only in a limited number of locations such as highly instrumented catchments. Across large regions, physically based numerical models have been shown to accurately describe transit time distributions when compared to tracer data, but these models often require extensive computational resources.

In this study, we examine machine learning approaches for efficient prediction of transit time, specifically investigating their spatial transferability across multiple large domains. We employ a continental scale physically based hydrologic model coupled with Lagrangian particle tracking to quantify transit time to baseflow metrics in four large river basins in the conterminous USA: Upper Colorado (290,000 sq km), Missouri (1,350,000 sq km), Upper Mississippi (490,000 sq km), and Ohio (420,000 sq km). We use results from the physically based model to train machine learning metamodels for predicting transit time metrics with multiple spatial aggregation units, with input predictors describing topography, climate, and geology. Functional input-output relationships learned by metamodels are assessed using model-agnostic explainability techniques and evaluated against theoretical physically based relationships. Spatial cross-validation frameworks are used to evaluate cross-domain predictive accuracy and characterize the influence of input data quantity and distribution similarity between training and target regions. Results from the analysis help elucidate the potential utility and limitations of machine learning metamodels for computationally efficient prediction of transit time metrics in data scarce regions.

How to cite: Soriano, M. and Maxwell, R.: Large-domain transferability of machine learning metamodels for predicting water transit time to baseflow, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12440, https://doi.org/10.5194/egusphere-egu26-12440, 2026.