Surface hydrologic connectivity is an important hydrologic characteristic for understanding hydrologic processes and catchment responses in depression-dominated areas. While tremendous recent interest remains focused on surface depression-induced hydrologic connectivity, few studies have illustrated the progressive development of hydro-connectedness at the watershed scale. The objective of this study is to identify and quantify threshold-controlled hydrologic connectivity behaviors for depression-dominated watersheds. We present a novel depression-oriented hydro-connectedness (HCd) characterization framework to recognize spatial distribution and topographic parameters of surface depressions and channels as well as the associated static/dynamic connections. The HCd was effectively applied to a depression-dominated watershed in North Dakota, U.S., and its unique capability was emphasized by the comparisons with other depression-oriented characterization algorithms/tools. The results provide an advanced understanding of the progressive expansion of connected areas and contributing area of the watershed outlet. The findings deliver new insights into threshold-controlled hydrologic connectivity and overland flow dynamics in depression-dominated areas, thereby facilitating water resource management and flood risk assessment, and supporting the development of depression-oriented hydrologic models for improved runoff prediction.
High-resolution precipitation data is essential for applications such as agriculture, infrastructure planning, and climate risk assessment, particularly in regions with complex terrain. Conventional climate datasets, including satellite and reanalysis products, are often too coarse to capture local variability. Downscaling techniques refine these data, but traditional dynamical methods are computationally intensive, and statistical methods require long-term observations that are often unavailable. We present a deep learning–based downscaling framework using transfer learning. A base model is pre-trained on large datasets and then fine-tuned to regions with limited local data, reducing the need for extensive weather station networks. Applied to Ecuador, the method produces accurate long-term precipitation estimates with only a few years of observations. This approach is flexible, extendable to other climate variables, and suitable for downscaling future climate projections, demonstrating the potential of transfer learning to improve the accessibility and quality of high-resolution climate information in data-scarce, topographically complex regions.
Environmental monitoring in complex rangeland ecosystems relies on human evaluation to navigate environment shifts. Local community knowledge is important for guiding herd mobility decisions within pastoral systems that use rangeland ecosystems. Currently, human-based scouting considers factors like rangeland location, season, and plant composition to make decisions based on traditional knowledge. However, climate change impact makes rangeland monitoring more complex by reducing the predictability of vegetation availability, species composition, access, and seasonal grazing patterns. To address this, there is a demand to automate rangeland assessment to help sustainable management. We present a novel human-augmented ecological assessment framework that formalizes local community knowledge into a scalable, knowledge-aware AI system through a multimodal Mixture of Experts (MoE) architecture with bidirectional cross-attention mechanism. This model is fed by a multimodal dataset from sub-Saharan Africa rangelands collected mainly via a mobile app where stakeholders encode decision-making parameters. This architecture helps the disambiguation of visual features by grounding them in seasonal metadata, which seasonal linguistic guides the model to prioritize site specific phenological patterns over undifferentiated spectral greenness. By employing a MoE layer to manage ecological data variability, our learning model uses community knowledge to override deceptive visual cues and improves rangeland quality classification over unimodal and standard fusion baselines. These results suggest that routing data through human-centric priors can improve rangeland monitoring.
Flood disasters are increasingly severe under global climate change, traditional flood risk assessment methods often overlook the interactions between natural and social factors, limiting their capacity to capture risk generation and transfer processes. To address this gap, this study integrates Morphological Spatial Pattern Analysis (MSPA) and Circuit Theory (CT), treating flood inundation areas as risk sources, analogizing the interaction between natural and social factors as a "Resistance network" and "Current" as a "Flood flow", and constructs a graded flood risk control network consisting of "Core risk sources-Flood corridors-Key nodes". Taking the July 2023 flood disaster in the Beijing–Tianjin–Hebei (BTH) region as a case study, the network identified 24 primary and 73 secondary risk sources, 253 transmission corridors, 327 pinch points, and 300 barrier points. Based on these findings, a chain-type prevention and control strategy of “Source control-Process interruption-Terminal regulation” was proposed.
Earth Surface System (ESS) open science data sharing platforms have been established globally. However, these platforms differ substantially in their access mechanisms, authentication procedures, query conventions, and front-end interactions, making cross-platform data acquisition difficult to automate through a single workflow. A structured survey of 26 representative ESS data platforms found that nine platforms (34.62%) provide neither a documented API nor a provider-supported SDK and rely exclusively on Web Portal access, while the available programmatic interfaces remain platform-specific. This paper designs and implements GeoNavigator, an agent-based intelligent download tool that provides a unified natural-language entry point for heterogeneous ESS data portals, regardless of whether programmatic access is available. GeoNavigator supports login, verification, querying, downloading, and local data validation within a single workflow. The Spatial-Aware DOM represents node semantics and viewport-level spatial features as structured observations for the LLM. Expert experience and distilled successful trajectories provide operational guidance and reduce repeated exploration. Across 84 retrieval tasks on seven ESS data platforms, GeoNavigator achieved a task success rate of 95.24% and a download accuracy of 94.05%. Tests with seven foundation models yielded success rates ranging from 70% to 90%, indicating compatibility with the evaluated model interfaces. These results demonstrate the feasibility of providing a common natural-language task interface across heterogeneous ESS portals, while the verified output files can be used in existing data-processing and modelling workflows.
The structural adequacy of classic infiltration models remains uncertain in non-stationary small watersheds. This study developed a diagnostic hybrid framework combining mechanism-enhanced infiltration models, Random Forest (RF)/Extreme Gradient Boosting (XGBoost) residual learning, and Shapley additive explanations (SHAP) interpretation. The framework was applied to 35 runoff events in the Xindiangou watershed on China’s Loess Plateau using modified Horton, Philip, and Green-Ampt models. RF/XGBoost residual learning improved calibration-period performance, but validation gains were model-dependent, indicating limited transferability. SHAP analysis identified model-specific residual signatures: Horton residuals were mainly linked to total rainfall, Philip residuals to maximum hourly rainfall intensity and vegetation conditions, and Green-Ampt residuals to antecedent wetness. Parameter–SHAP correlations suggested retrospective parameter compensation, especially for Green-Ampt. Bootstrap and collinearity diagnostics supported stable leading SHAP signals but required cautious group-level interpretation. Overall, explainable residual learning provides case-specific evidence to guide targeted refinement of infiltration models.
The Xinanjiang (XAJ) model is widely used and a de facto standard hydrological model in China. However, its implementation remains ambiguous due to the 5 mm inflow-based subdivision, originally introduced to control numerical errors in time discretization. Ambiguity persists regarding its necessity, scope, and threshold, leading to inconsistent implementations and hindering model reproducibility. To address this issue, we introduce a differential-form XAJ model and employ high-order numerical solvers to obtain a high-accuracy benchmark. Idealized experiment and large-sample applications are conducted to evaluate the subdivision method. Results show that subdivision is necessary for error control and should be applied across all modules. The 5 mm threshold is generally adequate, but smaller thresholds are required in a subset of watersheds. This study provides the first quantitative assessment of the subdivision method, resolves a long-standing ambiguity in the XAJ model, and highlights the importance of explicitly accounting for numerical implementation in hydrological modeling.
The high-sediment perched lower Yellow River forms a distinctive river-aquifer system, with groundwater decline along its banks over the past two decades. Using 2001–2020 data, a framework integrating Seasonal-Trend decomposition using Loess (STL) and an attention-enhanced Long Short-Term Memory (LSTM) network, SHapley Additive exPlanations (SHAP), and lag-response was constructed. STL indicated that long-term trends explained 53.12%–91.35% of groundwater-level variance, reflecting a trend-dominated decline. Model applicability varied with spatial and hydrogeological constraints: STL-LSTM resolved periodic near-river seepage (0–5 km; R2 = 0.888), whereas attention-enhanced LSTM improved predictions in pumped far-river zones (>10 km; R2 = 0.816). Driver attribution revealed spatial distance-decay in groundwater responses. Near-river groundwater dynamics (0–5 km) were mainly linked to river stage and runoff with an approximately 1-month lag, while high suspended sediment concentration may reduce river-aquifer exchange efficiency. Far-river dynamics (10–15 km) were dominated by agricultural water consumption (maximum contribution: −48.77%).
Connecting rivers and lakes can enhance regional water security, but artificial regulation may reshape ecological robustness. Focusing on the Feng River Basin, a mountain–plain transition zone, this study develops a directed topological network linking natural rivers and engineered canals and proposes a four-dimensional Structure–Connectivity–Function–Resilience (SCFR) framework. Driven by SWAT-based multi-scenario simulations, the framework quantifies the effects of drought stress and human intervention on composite water-network resilience. Results show that artificial systems expanded physical extent but reduced circularity, node connectivity rate, and overall connectivity by 11.83%, 3.04%, and 6.63%, respectively. Under extreme drought, ecological security rate and water system connection rate decreased by 7.68% and 1.68%, respectively, while recovery time increased by 0.622 days. Current water allocation measures did not fully offset resilience losses caused by artificial intervention. The SCFR framework developed in this study provides quantitative support for water system zoning, risk node identification, and connectivity regulation.
Large language model (LLM) agents are increasingly coupled with scientific simulators, yet how structured tool interfaces affect operational reliability and decision provenance remains unclear. This study introduces Hydro-MCP, a Model Context Protocol framework that constrains LLM-mediated SWAT + operation through typed function schemas, parameter-range validation, structured diagnostic feedback, and trajectory logging. We evaluated Hydro-MCP using a six-level ablation in the urban Jungnangcheon watershed, reference baselines, a rural Jiseokcheon stress test, and a held-out corruption task. In the permissive Jungnangcheon benchmark, the full Hydro-MCP configuration followed a shorter evaluation path within the tested workflow, 21.9 ± 5.3 versus 62.0 ± 16.5 model evaluations relative to unstructured LLM operation (p < 0.001), while terminal performance remained nearly unchanged (monthly NSE ≈ 0.893). A random-search baseline required fewer evaluations but achieved slightly lower terminal NSE, indicating that Hydro-MCP improves LLM-guided operation rather than non-LLM sampling efficiency. In Jiseokcheon, the full configuration achieved operational completion in all runs and Level 1 in 7 of 10 runs, whereas Levels 2-3 initiated no SWAT + simulation. This completion result concerns access to and execution of the workflow; the subsequent KGE, terminal parameters, and process-consistency analysis separately exposed zombie calibration despite acceptable fit. The corruption task further showed that logs and range constraints alone cannot detect physically corrupted model states. Within the tested workflow, the full Hydro-MCP configuration therefore improved operational control and produced reviewable provenance, while hydrological validity still required process-level consistency checks.
Urban stormwater systems generate large and heterogeneous datasets from in-situ sensors, remote sensing, and distributed monitoring networks, yet conventional models often struggle with nonlinear dynamics, data gaps, and scalability. This study presents a structured synthesis of machine learning (ML) approaches for urban stormwater modeling, focusing on their ability to address these limitations. We organize the literature across four domains — water quality prediction, flow modeling, infrastructure management, and real-time control — and introduce an integrated framework linking sensing, prediction, assessment, and adaptive management. Across the reviewed literature, ML provides the greatest value where stormwater applications involve heterogeneous data, nonlinear system behavior, and adaptive decision support; however, reliable deployment remains constrained by both technical and operational barriers including transferability, uncertainty, scalability, and workflow integration. We identify future research needs centered on hybrid physics–ML modeling, benchmark datasets, uncertainty-aware AI, digital twins, and transferable modeling frameworks to support robust and operational stormwater decision-support systems.
Monitoring inland and coastal water quality at large spatial scales and near-real-time frequency remains challenging because accessible cloud-based tools remain limited. We present Global-RSWQmonitor, an interactive Google Earth Engine (GEE) application for mapping and monitoring key water quality parameters worldwide. The platform integrates Landsat-8/9 and Sentinel-2 observations, incorporates ACOLITE-based atmospheric correction through a hybrid GEE workflow, and evaluates different atmospheric correction options for aquatic remote sensing. Chlorophyll-a, turbidity, total suspended solids, colored dissolved organic matter, and Secchi disk depth are retrieved using empirical models calibrated with the global GLORIA in-situ database. Validation against field measurements from representative lakes shows that the retrieved parameters capture major water quality patterns with reliable performance. By combining cloud computing, multi-sensor satellite archives, and a Python-based interface, Global-RSWQmonitor supports retrospective analysis, interactive visualization, and near-real-time water quality monitoring at global scales. The application is freely accessible at here.
Vapor intrusion is a dynamic process at hazardous waste sites where volatile contaminants in contaminated soil or groundwater vaporize and migrate through the soil strata and along preferential pathways into buildings where people can potentially be exposed. The paucity of data and the presence of other indoor sources of contaminants make assessment of the public health implications from vapor intrusion challenging, often requiring modeling and monitoring to develop multiple lines of evidence. This article describes and demonstrates a novel R package, vapintr, ATSDR created to facilitate the use of U.S. Environmental Protection Agency’s Johnson and Ettinger vapor intrusion model spreadsheet to allow the user to have the flexibility to evaluate multiple scenarios and enable stochastic simulation of multiple input parameters. This enables the user to understand the underlying sensitivity of the model to uncertain inputs. Additionally, vapintr aggregates and visualizes the inputs and outputs of the model alongside empirical site data and user supplied risk benchmarks and background values. Users can utilize the vapintr package using R code or through an interactive Shiny® application. To date, no R package has been created with the objective of facilitating evaluation of such sites using multiple lines of evidence. The vapintr package not only improves the efficiency of using the current Johnson and Ettinger Model spreadsheet but also provides enhancements in visualization, data management, and sensitivity analysis.
Planning support systems (PSS) based on geographic information systems (GIS) and multi-criteria decision analysis (MCDA) are sensitive to uncertainties in input parameters, yet these are often overlooked, with attention mainly on objective weighting. This study applies uncertainty analysis and Sobol global sensitivity analysis to assess key preference-related uncertainties in the Spatial Suitability Analysis Tool (SSANTO), a GIS-MCDA PSS for blue-green infrastructure (BGI) planning. We examined weights of objective, attribute value scales and the additivity parameter (γ) of the aggregation function, which controls how individual attributes are combined into overall suitability. Using spatially explicit Sobol analysis with 90,000 Monte Carlo simulations, we showed a limited subset of parameters drives outcomes: γ is dominant, followed by weights of objective, while value scales have minor effects. Spatial sensitivity is important, as low-average contributors may exert strong local effects. These findings guide parameter prioritization, data collection, and design of robust, transparent GIS-MCDA workflows, providing a framework for analyzing preference uncertainties.
Process-based hydrological models have advantages in representing physical mechanisms, whereas deep learning models excel at capturing complex nonlinear relationships. Effectively integrating these complementary strengths remains a key issue in runoff modeling. To address this issue, this study develops BiGRU-dXAJ, a differentiable hybrid hydrological model that places XAJ parameter learning within a multi-catchment joint training framework. The framework reformulates the XAJ model as differentiable tensor operations and couples it with a shared BiGRU-based parameter learning network. Meteorological forcings and static catchment attributes are jointly used to infer hydrological parameters, enabling end-to-end optimization of parameter learning and runoff simulation. Large-sample evaluation shows that dynamic-parameter BiGRU-dXAJ outperforms deep learning models, the traditional XAJ model, and static-parameter hybrid models. Spatial generalization experiments indicate promising but region-dependent performance in ungauged basins. The learned dynamic parameters generally follow the physical meanings defined in the XAJ model. Overall, differentiable hybrid modeling shows potential for runoff simulation.
Surface water storage (e.g., wetlands, lakes) is not typically considered in hydrological model calibrations. We tested a multivariate calibration process, incorporating Sentinel-1 and -2 surface water storage, for a Soil and Water Assessment Tool model across the 0.5 million km2 Upper Mississippi River Basin. While 19% of the 2000 parameter sets adequately simulated discharge (Kling-Gupta efficiency >0.5), only 5% also adequately simulated surface water storage (mean absolute error <2 m), reducing model output uncertainty. Using the best calibrated model, we found that changes in surface water storage capacity most strongly affected discharge during the first annual peak flow (i.e., floods), when storage was filling. Increases in upstream surface water storage capacity resulted in projected decreases in peak flow and flashiness, with changes persisting downstream to the watershed outlet. Our findings demonstrate the importance of surface water storage in multivariate model calibration processes to inform river discharge and flood impact predictions.
