Evangelos Rozos

The advent of machine learning (ML) has significantly transformed hydrology, particularly in the simulation of hydrological flows. However, ML techniques have not been employed to the same extent in stochastic hydrology. In applied sciences, the most common ML-based approach for developing stochastic simulation schemes is the use of generative adversarial networks (GANs), which consist of two sub-models, that is, a generator and a discriminator. Despite their potential, GANs have notable limitations, including high architectural complexity and the requirement to divide observed time series into shorter segments to generate sufficient training examples. This segmentation reduces the effective length of the series, limiting the model’s ability to capture and reproduce long-term dependencies. In this study, we propose a simpler stochastic scheme based on a single convolutional neural network (CNN) used as a generator, replacing the discriminator component of the GAN with a specifically designed cost function. The model is applied to a case study involving measured flow velocity time series and evaluated against traditional stochastic schemes designed for both Markovian and Hurst–Kolmogorov processes. Results show that the CNN-based approach not only offers computational simplicity but also outperforms conventional methods in preserving key statistical characteristics of the observed data.

The intensifying effects of climate change are increasingly disrupting the hydrological cycle (Barnett et al., 2005;Wu et al., 2013), contributing to more extreme droughts (Satoh et al., 2022), erratic precipitation and flow (O'Gorman et al., 2009;Diffenbaugh et al., 2017;Gudmundsson et al., 2021;Swain et al., 2025), and frequent flooding (e.g., Hirabayashi et al., 2013). These shifting patterns pose significant challenges for water resources planning and management, especially as traditional models often fall short in capturing the complex, nonlinear dynamics of climate-driven hydrologic systems (Milly et al., 2008). Accurate and timely forecasting of hydrological extremes has become more critical than ever-not only to reduce loss of life and property but also to guide long-term planning in water supply and ecosystem protection. However, the high variability and uncertainty associated with climate impacts demand approaches that can learn from data, adapt to changing conditions, and operate at finer spatial and temporal resolutions.Artificial intelligence (AI) offers a promising path forward-enabling the integration of vast and heterogeneous data sources, improving the precision of predictions, and supporting proactive decision-making (Kratzert et al., 2019;Camps-Valls et al., 2025). In particular, machine learning (ML) (Jordan and Mitchell, 2015) and deep learning (DL) (LeCun et al., 2015) models have shown great potential in extracting meaningful patterns from complex datasets such as remote sensing imagery, senor networks, reanalysis data, and hydrometeorological time series (Avand et al., 2021;Han et al., 2017;Nearing et al., 2024;Zhu et al., 2017). These AI techniques not only enhance predictive capabilities but also offer the flexibility to develop location-specific solutions, capture interactions between physical drivers, and improve lead times for early warning systems. Early contributions have laid the groundwork for AI applications in the field of hydrology and water resources. For instance, Kratzert et al. (2018) provided one of the first comprehensive applications of Long Short-Term Memory (LSTM), a type of DL model, in hydrology, demonstrating superior predictive skill over traditional models. Shen (2018) offered a broader perspective by reviewing DL applications across water sciences, highlighting how these models can not only enhance prediction but also serve as tools for scientific discovery by revealing underlying physical patterns. Complementing these efforts, Fang et al. (2017) employed DL to extend the Soil Moisture Active Passive satellite data, achieving seamless spatiotemporal coverage across the continental United States. Their approach significantly improved the resolution and continuity of soil moisture datasets, which are critical for hydrologic modeling and drought monitoring. Building on these advancements, Sawadekar et al. (2025) introduced a differentiable hydrologic modeling framework that fuses multiple precipitation datasets using interpretable neural networks. This fusion approach enhanced streamflow simulations, particularly for high-flow events, and provided spatially adaptive weighting of data sources, offering a more nuanced and accurate representation of precipitation inputs across diverse regions. Together, these foundational works underscore both the promise and the complexity of applying AI in hydrology and water resources, setting the stage for the more targeted and diverse studies included in this Research Topic. This Research Topiccollection brings together cutting-edge studies that apply AI methods to three central challenges in climate-impacted water systems: predicting precipitation and flood, downscaling coarse scale remote sensing water-related data, and developing interpretable deep learning models. The selected contributions span a range of modeling strategies-from LSTM-based neural networks and convolutional frameworks to ensemble machine learning techniques-each demonstrating how AI can be tailored to address specific needs while advancing scientific insight and practical application. A summary of each contribution is provided below.The contribution by Hamou-Ali et al. (2025) demonstrates the application of Random Forests (RF) for generating high-resolution Total Water Storage (TWS) maps. This approach integrates multiple remote sensing datasets, including precipitation (GPM, 10 km resolution), normalized difference vegetation index (NDVI, 1 km), land surface temperature (LST, 1 km), actual evapotranspiration (AET, 500 m), a digital elevation model (DEM, 30 m), and the normalized difference snow index (NDSI, 500 m). The initial RF output is a TWS map at 1 km resolution. To enhance its reliability, the model output is rectified using data from GRACE satellites, which provide TWS information at a coarser resolution of 100 km. Specifically, the RF output is first aggregated to match the GRACE resolution, and the residuals (differences from GRACE data) are then disaggregated back to 1 km. This correction step yields a more accurate TWS estimation, enhancing our understanding of groundwater resources-particularly crucial in arid regions where groundwater plays a vital role in drought resilience. 2025) explore the use of six ML models-Decision Tree, Random Forest, K-Nearest Neighbors (KNN), AdaBoost, XGBoost, and Long Short-Term Memory (LSTM)-for monthly precipitation forecasting. After comparing the performance of individual models, the study proposes a two-layer stacked learning approach. The first layer is trained on the original data, while the second layer, a meta-learner, is trained on the output of the first. This ensemble method achieved outstanding results, reducing the Root Mean Squared Error by more than 50% compared to the best-performing single model. Such improvement is particularly valuable in arid regions affected by the El Niño Southern Oscillation (ENSO), where climate complexity challenges forecasting efforts,...

Amid the growing energy–water nexus crisis, large dams are being reconsidered as viable solutions despite significant environmental concerns. A critical and enduring issue with large dams is the threat they pose to downstream communities and infrastructure in the event of structural failure. The Oroville Dam spillway incident, where inadequate maintenance led to uplift forces that exceeded the structural capacity of a chute slab, causing severe damage, has renewed the focus on the structural stability of spillway components. This study argues that conventional methods, which rely on averaged values and empirical coefficients, may be inadequate for accurately capturing the dynamical stresses on spillway chutes induced by turbulent flow conditions. We propose a novel approach using stochastic simulation schemes to generate synthetic time series of velocity, which are then applied to a differential equation governing the chute slab oscillations. Through a hypothetical case study inspired by the Oroville incident, we demonstrate two key issues: first, that the conventional approach significantly underestimates the maximum stresses experienced by chute slabs under dynamic uplift pressures; and second, that the stochastic structure of the velocity, particularly the variance and persistence, plays a major role in determining the maximum stress.

Assessing drag force and Reynolds stresses in turbulent flows is crucial for evaluating the stability and longevity of hydraulic structures. Yet, this task is challenging due to the complex nature of turbulent flows. To address this, physical models are often employed. Nonetheless, this practice is associated with difficulties, especially in the case of high sampling frequency where the inherent randomness of velocity fluctuations becomes mixed with the measurement noise. This study introduces a stochastic approach, which aims to mitigate bias from measurement errors and provide a probabilistic estimate of extreme stress values. To accomplish this, a simple experimental setup with a hydraulic jump was employed to acquire long-duration velocity measurements. Subsequently, a modified first-order autoregressive model was applied through ensemble simulations, demonstrating the benefits of the stochastic approach. The analysis highlights its effectiveness in estimating the uncertainty of extreme events frequency and minimizing the bias induced by the noise in the high-magnitude velocity measurements and by the limited length of observations. These findings contribute to advancing our understanding of turbulent flow analysis and have implications for the design and assessment of hydraulic structures.

Storm Daniel initiated on 3 September 2023, over the Northeastern Aegean Sea, causing extreme rainfall levels for the following four days, reaching an average of about 360 mm over the Peneus basin, in Thessaly, Central Greece. This event led to extensive floods, with 17 human lives lost and devastating environmental and economic impacts. The automatic water-monitoring network of the HIMIOFoTS National Research Infrastructure captured the evolution of the phenomenon and the relevant hydrometeorological (rainfall, water stage, and discharge) measurements were used to analyse the event’s characteristics. The results indicate that the average rainfall’s return period was up to 150 years, the peak flow close to the river mouth reached approximately 1950 m³/s, and the outflow volume of water to the sea was 1670 hm³. The analysis of the observed hydrographs across Peneus also provided useful lessons from the flood-engineering perspective regarding key modelling assumptions and the role of upstream retentions. Therefore, extending and supporting the operation of the HIMIOFoTS infrastructure is crucial to assist responsible authorities and local communities in reducing potential damages and increasing the socioeconomic resilience to natural disasters, as well as to improve the existing knowledge with respect to extreme flood-simulation approaches.

Highly turbulent water flows, often encountered near human constructions like bridge piers, spillways, and weirs, display intricate dynamics characterized by the formation of eddies and vortices. These formations, varying in sizes and lifespans, significantly influence the distribution of fluid velocities within the flow. Subsequently, the rapid velocity fluctuations in highly turbulent flows lead to elevated shear and normal stress levels. For this reason, to meticulously study these dynamics, more often than not, physical modeling is employed for studying the impact of turbulent flows on the stability and longevity of nearby structures. Despite the effectiveness of physical modeling, various monitoring challenges arise, including flow disruption, the necessity for concurrent gauging at multiple locations, and the duration of measurements. Addressing these challenges, image velocimetry emerges as an ideal method in fluid mechanics, particularly for studying turbulent flows. To account for measurement duration, a probabilistic approach utilizing a probability density function (PDF) is suggested to mitigate uncertainty in estimated average and maximum values. However, it becomes evident that deriving the PDF is not straightforward for all turbulence-induced stresses. In response, this study proposes a novel approach by combining image velocimetry with a stochastic model to provide a generic yet accurate description of flow dynamics in such applications. This integration enables an approach based on the probability of failure, facilitating a more comprehensive analysis of turbulent flows. Such an approach is essential for estimating both short- and long-term stresses on hydraulic constructions under assessment.

Machine learning has been used in hydrological applications for decades. Recent studies, after a systematic comparison, have shown that machine learning models (more precisely, deep learning with thousands of nodes) can outperform even the most sophisticated physically based models. Furthermore, one of the basic criticisms, that machine learning produces black box models, has been addressed by researchers, who have indicated how this black box can be made transparent to obtain explainable/interpretable results. However, the main disadvantage of the machine learning approaches (especially deep learning, which may employ hundreds of thousands of parameters) remains the CPU-intensive training process. This disadvantage can be overcome by employing hybrid modelling frameworks that combine simple machine learning models with parsimonious hydrological models. The drawback of these parsimonious approaches is the susceptibility of the latter to conditional systematic errors, which propagate through the modelling framework and cannot be eliminated by simple machine learning networks (employing complex networks would nullify the sought benefit of reduced CPU times). In this study, we suggest methods to cope with this kind of error and achieve a modelling performance close to the best achievable with the available data.

Machine learning has been used in hydrological applications for decades, and recently, it was proven to be more efficient than sophisticated physically based modelling techniques. In addition, it has been used in hybrid frameworks that combine hydrological and machine learning models. The concept behind the latter is the use of machine learning as a filter that advances the performance of the hydrological model. In this study, we employed such a hybrid approach but with a different perspective and objective. Machine learning was used as a tool for analyzing the error of hydrological models in an effort to understand the source and the attributes of systematic modelling errors. Three hydrological models were applied to three different case studies. The results of these models were analyzed with a recurrent neural network and with the k-nearest neighbours algorithm. Most of the systematic errors were detected, but certain types of errors, including conditional systematic errors, passed unnoticed, leading to an overestimation of the confidence of some erroneously simulated values. This is an issue that needs to be considered when using machine learning as a filter in hybrid networks. The effect of conditional systematic errors can be reduced by naively combining the simulations (mean values) of two or more hydrological models. This simple technique reduces the magnitude of conditional systematic errors and makes them more discoverable to machine learning models.

This paper describes HYDRONET, a telemetry-based prototype of a streamflow monitoring network in the Greek territory, where such data are sparse. HYDRONET provides free and near-real-time online access to data. Instead of commercially available stations, in-house-designed and -built telemetric stations were installed, which reduced the equipment cost by approximately 50%. The labour of hydrometric campaigns was reduced by applying a new maximum-entropy method to estimate the discharge from surface velocity observations. Here, we describe these novelty elements succinctly. The potential of HYDRONET to provide civil protection services is exemplified by a flood warning demonstrator for Kalamata’s City Centre. The network’s operation, including the hydraulic criteria for monitoring site selection, the characteristics of the telemetric equipment, the operational monitoring and hydrometric procedures, and the specifics of data transmission, quality control, and storage are described in detail, along with experiences with problems encountered during this pilot phase.