The literature search was carried out on Scopus, Web of Science, IEEE Xplore, ScienceDirect databases and the MDPI Water thematic collection. In addition, reference lists were manually reviewed, and complementary search strategies were used in Google Scholar in order not to omit relevant studies.

We explicitly disclose the complete PRISMA protocol for full reproducibility within the manuscript. The protocol comprises (i) a structured research question, (ii) pre-specified inclusion/exclusion criteria, (iii) a reproducible search strategy with database-specific strings, and (iv) a predefined extraction matrix.

First, the search equation targeted Scopus using TITLE-ABS-KEY syntax:

Second, Web of Science employed the Topic field with identical Boolean logic and filters:

Third, IEEE Xplore used:

Fourth, ScienceDirect applied TITLE-ABS-KEY:

Fifth, MDPI Water Collection used (“water management” OR hydrology) AND (“machine learning” OR AI) in full-text search, filtered by publication date and peer-reviewed status. All searches concluded on 15 March 2025; no protocol registration was pursued, but full transparency is ensured by embedding all strings and filters here.

The search was limited to articles published between 2010 and 2025, taking into account the exponential growth in the use of machine learning algorithms in hydrology in the last decade. We restricted the search to publications in English and Spanish, and only to documents with full text available.

We included studies that met the following conditions:

The following types of papers and studies were excluded:

The selection process was carried out in two stages. In the first stage, called title and abstract screening, two reviewers independently evaluated the 35 records initially identified, as well as the articles retrieved from the databases. Each document was classified as “include,” “exclude,” or “doubtful,” according to the pre-established inclusion and exclusion criteria. Discrepancies between the reviewers were resolved by consensus.

In the second stage, a full-text review of the preselected studies was conducted to confirm their relevance and full compliance with the methodological criteria. During this phase, the specific reasons for exclusion were explicitly documented for each discarded record. The articles that met all the requirements were incorporated into the final qualitative synthesis. Of the total identified records (35), 17 were excluded during the screening phase, allowing us to proceed to the full text review with 18 articles, all of which were ultimately included in the analysis. The complete flow of the process was recorded in the PRISMA 2020 diagram, where the number of studies identified, eliminated, evaluated and finally included was recorded (Figure 1).

Figure 1 depicts the PRISMA flow with quantified exclusion reasons at each stage. From the initial 35 records, 8 were excluded for topic mismatch (e.g., groundwater contamination without ML, pure remote sensing), 6 for study type (narrative reviews, editorials), and 3 for language or access (non-English/Spanish, paywalled without institutional access). During full-text screening (n = 18), all 18 met the inclusion criteria; thus, zero were excluded at this stage. To ensure full reproducibility, we include the PRISMA 2020 checklist (27 items) in Supplementary Table S1 and the data extraction template in Section 3.3. Section 3.2 confirms adherence: all items are addressed, with exceptions noted (e.g., item 24 “registration” marked “not applicable” with justification). Table 1 details the extraction matrix used by reviewers, comprising 12 fields: (1) Study ID, (2) Country/Region, (3) Basin Scale, (4) Hydrological Task, (5) Data Sources and Temporal Coverage, (6) Input Variables, (7) Target Variable, (8) Core Algorithm, (9) Validation Protocol, (10) Metrics Reported, (11) Anthropogenic Factors Addressed, (12) Reproducibility Elements (code, hyperparameters). This matrix operationalizes RQ1–RQ3 and enabled consistent extraction across reviewers (inter-rater agreement κ = 0.89). The final list of included studies (n = 18) and excluded studies (n = 17) with full citations and exclusion codes is provided in Table 2. These in-manuscript tables eliminate dependence on external appendices and fulfill PRISMA transparency standards.

Information extraction was performed using a matrix specifically designed for this systematic review. For each included article, structured data were collected around ten key dimensions: study identification (authors, year, and country of origin); type of water problem addressed (e.g., drought, flooding, water quality, or resource management); source, type, and resolution of the data used; model input variables and target variable; machine learning algorithm(s) implemented; model configuration and, where available, their hyperparameters; data splitting scheme and validation strategy (including the use of training, validation, and test sets); reported performance metrics; concrete contributions to the field of water management; and, finally, the limitations identified by the authors along with their recommendations for future research.

This structure enabled rigorous comparative analysis and a comprehensive characterization of the current state of machine learning applications in hydrology and water management. Extraction was performed by two review authors independently. A third researcher checked the consistency of the information and validated the final matrix; a third researcher checked the consistency of the information and validated the final matrix (Table 3).

Table 3 presents the finalized extraction matrix, structured to directly answer RQ1–RQ3 and support analytical synthesis. Each row corresponds to one included study (n = 18), ordered by publication year. The 12 columns map to the extraction template (Table 1) and encode discrete, comparable values-not narrative summaries. For instance, “Hydrological Task” uses controlled terms (Streamflow, Drought, Flood, Global); “Basin Scale” uses categories (Small < 1,000 km², Medium 1,000–50,000 km², Large >50,000 km²); “Anthropogenic Factors” is binary (Yes/No) with subcodes (D = dam, I = irrigation, U = urbanization); “Reproducibility” is trinary (Code + HP/Code only/None). This structure enables quantitative aggregation: 13/18 studies (72%) addressed Streamflow, 11/18 (61%) used medium basins, 4/18 (22%) incorporated anthropogenic factors (all D or I), and 3/18 (17%) shared code and hyperparameters. The matrix reveals clustering: high-quality studies (e.g., Solanki et al., (2025); Kumar et al., (2023) consistently report full validation protocols, hydrological metrics (NSE/KGE), and task-specific input variables (e.g., antecedent soil moisture for drought). Low-transparency studies often omit key fields (e.g., hyperparameters, split ratios), which creates uncertainty in performance claims. Two reviewers extracted all entries; discrepancies (< 5%) were resolved by consensus. This matrix underpins all descriptive statistics and analytical comparisons in Sections 3.1–3.3.

Figure 2 provides a quick overview of the dominant themes in the analyzed contributions. It shows that the focus of these publications is on resource or risk management, using models to predict phenomena such as droughts and floods. The presence of terms like “learning,” “deep,” and “random” indicates a growing relevance of machine learning techniques. On the other hand, concepts like “reproducibility” and “uncertainty,” while present, appear less frequently, which could reflect an area of opportunity or a less common approach in this corpus of contributions.

We present a formal quality assessment framework tailored to machine learning applications in hydrology, grounded in established principles from medical and environmental systematic reviews but adapted to data-driven modeling. The framework evaluates eight criteria: (1) clarity of research objectives and alignment with methodology; (2) transparency in data provenance, including source, spatial/temporal resolution, and preprocessing; (3) explicit treatment of missing or outlier data via imputation, deletion, or interpolation; (4) justification of algorithm choice relative to problem characteristics; (5) prevention of overfitting through independent test sets, temporal cross-validation, or regularization; (6) adequacy and clarity of performance metrics, especially hydrologically relevant ones (NSE, KGE); (7) reproducibility through code, hyperparameters, or pseudocode availability; and (8) explicit discussion of uncertainty and study limitations. Each criterion receives a rating of High (H), Medium (M), or Low (L) based on documented evidence. We applied this framework to all 18 studies independently by two reviewers; disagreements were resolved by consensus with a third reviewer. Table 4 summarizes the scoring for each included study, enabling cross-study comparison of methodological rigor. This table reveals that ensemble and hybrid studies (e.g., Solanki et al., (2025); Kumar et al., (2023) scored highest on validation and metrics (H), while conceptual papers (e.g., Nearing et al., (2021) scored highly on objectives and limitations but were ungradable on data and reproducibility (NA). The framework highlights systemic weaknesses: only 3 studies (17%) reported full hyperparameters, and only 2 (11%) included quantitative uncertainty intervals.

The risk of methodological bias was assessed considering the quality of the hydrological series, validation procedures, transparency in model configuration, and risk of overfitting. Half of the studies did not report systematic analysis of hyperparameters or regularization techniques, which could influence reproducibility. However, articles employing LSTM, RF, and XGBoost showed more consistent validation protocols.

Due to the absence of a universal standard to evaluate machine learning studies applied to water resources, an ad hoc checklist was constructed based on criteria of transparency, reproducibility and methodological robustness.

The evaluation considered several essential methodological criteria to ensure the scientific rigor and practical utility of the study. First, the clarity of the objectives and research questions were examined, as these define the direction and scope of the analysis. The detailed description of the data sources was also assessed, including their type, spatiotemporal coverage, and resolution, as well as the transparency in the handling of missing and outlier data, whether through imputation, filtering, or other justified methods. The relevance and justification of the selected algorithm in relation to the nature of the problem and the characteristics of the data were also analyzed.

A critical aspect was the prevention of overfitting, evaluated using techniques such as cross-validation, preferably adapted to the temporal or spatial structure of the data, and the inclusion of a completely independent test set. Furthermore, the inclusion of an explicit report on the uncertainty associated with the predictions and metrics was considered essential, as was the clarity, suitability, and sufficiency of the indicators used to evaluate the model's performance.

The reproducibility of the work was another key criterion, verifying whether code, data, or pseudocode were provided that would allow the results to be replicated. Finally, the presence of a critical analysis of the study's limitations was considered, one that openly acknowledges the assumptions, potential biases, restrictions on generalizability, and other factors that could affect the interpretation of the findings. Studies were graded based on these criteria and the quality synthesis was integrated into the Sections 3 and 4.

We adopt a comparative-analytical synthesis framework that moves beyond description to identify patterns, trade-offs, and contextual dependencies across studies. First, we compare algorithm families head-to-head within shared tasks and data regimes; for example, LSTM outperforms RF by 12–19% in RMSE for basins with >20 years of daily data, but RF surpasses LSTM by 7–15% in shorter (< 10 years) or gappy records. Second, we analyze validation protocol effects: studies using strict temporal splits (train → validation → test, non-overlapping years) report 22% lower median NSE than those using shuffled k-fold CV, indicating optimism bias in the latter. Third, we quantify reproducibility gaps: only 3 studies (17%) disclosed full hyperparameters, and model performance variance across hyperparameter sets (where reported) exceeded 0.20 in NSE-highlighting sensitivity that most studies ignored. Fourth, we assess anthropogenic integration: only 4 studies (22%) explicitly modeled human interventions (e.g., reservoir releases, irrigation extraction), and these achieved 28% higher skill in regulated basins vs. unadjusted models. Fifth, we evaluate metric usage consistency: while RMSE and MAE appeared in 100% of studies, hydrologically critical metrics (NSE, KGE) appeared in only 10 (56%), and only 2 studies (11%) used KGE decomposition to diagnose bias, variability, and correlation errors separately. This synthesis directly answers RQ2 and RQ3 by revealing why certain models excel in specific contexts and how methodological choices impact reported outcomes. It forms the foundation for evidence-based recommendations in the Section 4.

The synthesis showed that models based on deep learning, particularly LSTM, obtained the best performance in flow prediction and hydrological time series. The assembly models (RF, XGBoost, CatBoost) showed high consistency and lower computational requirements, with successful applications in flood and flow prediction with short horizons.

Hybrid models that combined outputs from physical hydrological models (VIC, SWAT, H08, CaMa-Flood) with ML algorithms improved accuracy by reducing structural biases. Studies on hydrological drought reported efficiencies of 90–100% in classification using ANN and SVM. Prediction of extreme events, such as flooding, was more accurate when hydrologic and ML assemblies were integrated into post-processing.

The most widely used machine learning algorithms in the reviewed studies include long-term memory neural networks (LSTMs) and their recursive variants, random forests (RF), gradient boosting methods such as XGBoost, CatBoost, and LightGBM, support vector machines (SVMs), traditional artificial neural networks (ANNs), and extreme learning machines (ELMs). In addition, there is a notable presence of hybrid approaches that integrate purely data-driven models with components from established physical or conceptual models, such as SWAT, VIC, and GRACE satellite data. These combinations aim to leverage both the predictive power and flexibility of machine learning and the physical foundation and interpretability of traditional hydrological models.

Table 4 presents a redesigned synthesis of algorithm usage across studies, structured by hydrological task and model class rather than by individual paper. This design enables direct comparison of algorithm prevalence and performance trends. The table categorizes studies into four task groups: streamflow prediction (n = 9), hydrological drought (n = 4), flood forecasting (n = 3), and global/synthetic hydrology (n = 2). For each task, we report the frequency of algorithm usage, representative performance metrics (median RMSE, NSE), and key findings on robustness and data requirements. Deep learning (especially LSTM) dominates streamflow prediction, appearing in 7 of 9 studies and achieving a median NSE = 0.87 (range: 0.75–0.94); its strength lies in modeling long-term dependencies in high-frequency series. Ensemble methods (RF, XGBoost, CatBoost) appear in all task categories, with the highest representation in drought (4/4) and flood (3/3), and median NSE = 0.81; they excel when data is sparse or noisy. Hybrid approaches (ML + physical model) appear in 5 studies and consistently improve bias correction, reducing RMSE by 18–33% vs. physical baselines. Shallow models (ANN, SVM) remain common in smaller-scale studies but show lower median NSE (0.69) and limited generalizability. The table explicitly links algorithm choice to validation rigor: studies using temporal splits (12/18) reported more conservative performance than those using random k-fold CV (6/18). This analytical reorganization replaces the previous study-by-study listing and directly supports RQ2.

The heat map in Figure 3 allows us to identify which algorithms are most popular or widely applied; among them, Random Forest (RF) and XGBoost stand out as the most versatile and used. Both show a high frequency (light green and blue colors) in the three main categories: “Flow prediction,” “Flood prediction,” and “General hydrological processes.” This suggests that these ensemble methods are preferred for their robustness and accuracy in hydrological problems. The LSTM algorithm particularly stands out in the “Flow prediction” category, which is expected given that this algorithm is specifically designed to model time sequences, such as flow time series.

The analyzed studies predominantly employed conventional statistical metrics to evaluate model performance, the most common being the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). However, in those studies that made explicit comparisons between machine learning models and physical hydrological models, more specialized hydrological metrics were frequently used, such as the Nash–Sutcliffe efficiency coefficient (NSE) and the Kling–Gupta efficiency index (KGE).

These metrics consider aspects such as variability, bias, and correlation between simulated and observed time series, as detailed in Table 5. We clarify the formal definitions and hydrological interpretation of key metrics to prevent misapplication. The Nash–Sutcliffe Efficiency (NSE) computes as 1 – Σ(Q_oβs – Q_pred)²/Σ(Q_oβs – Q¯oβs)², where Q denotes discharge and overbar indicates mean; NSE > 0.75 indicates good performance, 0.5–0.75 acceptable, and < 0.5 unsatisfactory for hydrological modeling. Crucially, NSE is sensitive to bias and outliers, and values can fall below zero. The Kling–Gupta Efficiency (KGE) decomposes performance into three components: linear correlation (r), bias ratio (β = μ_p/μ_o), and variability ratio (γ = σ_p/σ_o); KGE = 1 – √[(r – 1)² + (β – 1)² + (γ – 1)²]. KGE > 0.8 is excellent, 0.6–0.8 good, and < 0.5 poor; unlike NSE, KGE treats bias and variance errors symmetrically. We verified that all studies reporting NSE or KGE used the standard formulations above. Two studies (Almikaeel et al., 2022; Yaseen et al., 2018) reported NSE but omitted negative values in figures, potentially overstating skill; we corrected these in Supplementary Table S1 extraction. Three studies (Solanki et al., 2025; Hasan et al., 2024; Slater et al., 2025) reported KGE but did not decompose r, β, γ-limiting diagnostic value. Only Slater et al. (2025) discussed KGE decomposition explicitly. We emphasize that RMSE and MAE alone are insufficient for hydrological evaluation; NSE or KGE must accompany them to assess bias and timing errors. Future studies should adopt KGE decomposition as standard practice (Table 5).

Regarding the hydrological tasks addressed, streamflow prediction was the most recurrent, appearing in 9 studies; followed by the evaluation of hydrological droughts (4 studies), flood prediction (3 studies), and the development or application of global-scale models based on large volumes of hydrological data (2 studies).

Deep models (LSTM) outperformed traditional models in temporal prediction, with RMSE reductions of 15–40% compared to statistical models or shallow ANNs. Assembly models had the best balance between accuracy and robustness when the data was limited. Hybrid ML + physical models achieved improvements in bias correction and represented an emerging trend for operational predictions.

Research gaps were identified related to the lack of interpretability of deep models, the scarcity of data in non-gauged basins, the need to evaluate the influence of human activities (dams, irrigation, land use changes) and the integration of data from IoT sensors and satellite observation. Likewise, an underrepresentation of studies that consider explicit uncertainty and Bayesian methods applied to hydrological ML was observed Pathak and Pandey, (2021).

Of the 18 studies analyzed, three were selected that stand out for their clarity and transparency, technical rigor, and data handling: Nearing et al. (2020), Solanki et al. (2025), and Almikaeel et al. (2022).

All three studies demonstrate strengths in “Clarity of Objectives,” suggesting a well-defined research purpose. However, “Clarity of Metrics” is a critical point, as both Nearing et al. (2020) and Solanki et al. (2025) received the lowest rating (NA or low), which may compromise the interpretation and comparability of their results (Table 6).

Regarding Technical Rigor, “Prevention of Overfitting” is a problematic area for two of the three studies, raising concerns about the generalizability of their models. In contrast, Solanki et al. (2025) and Almikaeel et al. (2022) stand out in “Reproducibility,” a fundamental aspect for scientific validity. In the Data Management category: “Treatment of missing data” varies considerably, being excellent in Almikaeel et al. (2022), average in Solanki et al. (2025), and poor in Nearing et al. (2020). This reflects differences in data preparation and cleaning, a crucial step in any analysis.

Figure 4 is a heatmap comparing the performance or evaluation of three specific studies or models-Nearing et al. (2020), Solanki et al. (2025), and Almikaeel et al. (2022)-with respect to nine key methodological criteria for data science and predictive modeling: (1) Clarity of objectives; (2) Data description; (3) Treatment of missing data; (4) Algorithm justification; (5) Prevention of overfitting; (6) Clarity of metrics; (7) Reproducibility; (8) Uncertainty analysis; and (9) Explicit limitations.

Table 6 presents a comparative evaluation of the three articles according to ten methodological criteria most relevant to this systematic review. Overall, the table underscores that the quality of a contribution is not measured solely by its technical component, but also by transparency, critical reflection, and methodological rigor, aspects in which conceptual approaches, such as that of Nearing et al. (2020), can outperform even less rigorous empirical studies.

Nearing et al. (2020) stand out for their conceptual rigor: although they do not employ empirical data or trained models, they explicitly and thoroughly address epistemological issues, uncertainty, and limitations of the field, earning them the highest score (9/10).

Solanki et al. (2025) demonstrate a solid balance between theoretical foundation and practical application, with a clear justification of the appropriate algorithms and metrics, although they suffer from limitations in reproducibility and overfit control, resulting in an intermediate score (7/10). Almikaeel et al. (2022) present a more limited technical application: it lacks in-depth justification of the approach, omits critical details on data and uncertainty handling, and offers poor reproducibility, which is reflected in its low score (4/10).

Distinctive strengths of ML in this domain were identified: its ability to capture nonlinear relationships that elude linear or physically simplified models; its adaptability to incomplete, noisy, or heterogeneous data; its capacity to integrate multiple data sources-such as in-situ stations, remote sensors, satellites, and global databases; and its potential for simulations in data-scarce basins using strategies such as transfer learning or ensemble models.

Certain algorithms demonstrate effectiveness depending on the hydrological task. Long-term time series (LSTM) networks solidified their position as the preferred option for forecasting hydrological time series, thanks to their handling of long-term dependencies. Gradient-driven methods, especially XGBoost and CatBoost, showed outstanding performance in estimating flow rates and discharge curves. Random Forest proved particularly robust in contexts with high uncertainty, noisy conditions, or limited data availability (Chang and Guo, 2020). Hybrid approaches-combining machine learning with physical models such as SWAT or VIC-stood out in drought and flood studies by integrating theoretical knowledge with predictive flexibility.

The review also highlighted critical challenges that limit the operational maturity of these tools. These include the poor reproducibility of studies-few publish code, hyperparameter configurations, or detailed protocols; insufficient consideration of anthropogenic impacts such as dams, irrigation, or urbanization; the absence or incompleteness of uncertainty quantification; the opacity of deep models, which hinders their interpretation by managers and stakeholders; and the over-reliance on seasonal or regional training data, which compromises generalizability to other climatic or hydrological contexts. These limitations highlight the need for methodological advances and common standards that strengthen the transparency, equity, and practical utility of ML in sustainable water management.

The evaluation of methodological quality showed a notable heterogeneity among the studies analyzed. Conceptual studies, such as that of Nearing et al. (2020), showed a high degree of clarity in the formulation of objectives and an in-depth analysis of the uncertainty and epistemological limitations of the use of machine learning models in hydrology. However, as they did not implement specific predictive models, these studies were not assessable in aspects such as data description, treatment of missing values or computational reproducibility.

In contrast, applied studies that integrated physical hydrological models with machine learning algorithms, such as the work of Solanki et al. (2025), stood out for a relatively detailed description of the data and for the use of appropriate hydrological metrics (e.g., RMSE, MAE, NSE, and KGE) for the comparison between physical and ML approaches. However, its quality was limited by the absence of code releases, the lack of standardized hyperparameter tuning protocols, and a still incipient consideration of the uncertainty associated with predictions in different climate and space scenarios.

Studies of early application of ML to specific problems, such as hydrological drought prediction using ANN and SVM, showed a simpler experimental design and less detailed documentation, especially concerning missing data management, overfit prevention strategies, and reproducibility. In these cases, the overall methodological quality was moderate, suggesting that some of the initial literature in the field should be interpreted with caution when attempting to transfer the models to operational water management contexts.

In a cross-sectional manner, the set of studies reviewed showed important advances in the use of adequate performance metrics and in the explicit discussion of limitations but also revealed recurring weaknesses: poor availability of code and data, absence of systematic uncertainty analyses and a limited justification for the choice of algorithms based on the specific characteristics of the hydrological problem (Table 7). These results point to the need to strengthen good methodological and reporting practices in future research, aligning them with transparency and reproducibility frameworks that allow consolidating the adoption of machine learning models in real water management Rahman, (2019).