Advances in machine learning for agricultural water management: a review of techniques and applications

ART

adaptive resonance theory

BWET

bias-corrected weighted ensemble technique

GWFP

groundwater flow prediction

BWFP

baseflow prediction

CNN

convolutional neural network

DAT

data assimilation technique

DT

decision tree

ELM

extreme learning machine

GAM

generalized additive model

GBR

gradient boosting regressor

GWET

groundwater evapotranspiration

KNN

K-nearest neighbors

LR

linear regression

LSTM

long Short-Term Memory

MAE

mean absolute error

MAPE

mean absolute percentage error

MLP

multi-layer perceptron

RAE

relative absolute Error

RFR

random forest regressor

RL

reinforcement learning

RRSE

root relative squared error

RSS

residual sum of squares

SVR

support vector regression

SWAT

soil and water assessment tool

WF

wavelet function

To enhance crop water productivity (CWP), two main strategies can be employed: increasing crop yields without raising water consumption or reducing water usage while sustaining or improving yields (Foley et al. 2020). On a global scale, improving CWP requires a thorough understanding of water usage patterns, including where and how water is utilized and its variability in producing specific crop quantities (Blatchford et al. 2018; Kilemo 2022). In recent years, the intersection of machine learning (ML) techniques and agricultural water management has garnered significant attention due to its potential to address pressing challenges in sustainable food production. The increasing strain on water resources due to population growth, climate variability, and competing sectoral demands makes optimizing agricultural water usage critical for food security and sustainability (Hameed et al. 2019; Jung et al. 2021). ML provides a data-driven approach to addressing these challenges by developing predictive models and optimizing resource management (Sarker 2021; Umutoni & Samadi 2024). Advanced algorithms such as neural networks, decision trees, and support vector machines (SVM) enable researchers to analyze water use patterns and develop innovative solutions for mitigating water-related risks (Veeragandham & Santhi 2020).

Water is a fundamental resource in agriculture, essential for crop growth, livestock production, and ecosystem health. The efficient and sustainable use of water is critical to ensure food security, support rural livelihoods, and preserve natural ecosystems. However, water scarcity, pollution, and competition from other sectors pose significant challenges to agricultural water management (D'Odorico et al. 2020). In many regions, unsustainable water practices, such as excessive irrigation and groundwater depletion, threaten the long-term viability of agricultural systems. Addressing these challenges requires innovative approaches that optimize water usage, improve water productivity, and promote conservation practices that minimize environmental impacts (Kılıç 2020).

While ML offers transformative potential, its application in agriculture faces significant hurdles, including the need for high-quality data, model interpretability, and integration with existing practices. This review addresses these challenges by showcasing original achievements, such as the superior performance of ensemble methods (e.g., RF, light gradient boosting machine (LightGBM)) in predicting water footprints and evapotranspiration, and the integration of real-time data streams to improve decision-making. These advancements demonstrate how ML can overcome traditional limitations, offering robust solutions for sustainable water management.

ML is a subfield of artificial intelligence (AI) that focuses on developing algorithms and statistical models that enable computers to learn from and make predictions or decisions based on data (Lee et al. 2017). In the context of agricultural water management, ML plays a crucial role in analyzing vast datasets related to water usage patterns, soil characteristics, climate variables, and crop growth parameters (Sahoo et al. 2017; Benos et al. 2021). By leveraging ML techniques, such as pattern recognition and predictive modeling, researchers can gain valuable insights into the complex interactions between these factors and optimize water resource allocation for sustainable agricultural practices. Recent advancements in ML have introduced sophisticated models such as EMD-LSTM, LSTM-INFO, RVFL-EROA, ANFIS-WCAMFO, and ANN-RUNAO, which show promise in time series modeling for hydrological and environmental applications (Li et al. 2018; Ikram et al. 2023; Adnan et al. 2024, 2025). However, their application in agricultural water management remains limited due to a lack of validation studies and practical implementation in this domain.

However, despite the growing interest and potential benefits, the application of ML in agricultural water management presents several challenges and opportunities that warrant further investigation (Li et al. 2023). These include the need for robust data collection and quality assurance protocols, the development of interpretable and transparent models, and the integration of socio-economic factors and stakeholder perspectives into decision-support frameworks. Moreover, as climate change exacerbates water scarcity and variability, there is a pressing need to enhance the resilience of agricultural systems through adaptive management strategies informed by advanced analytics (Maleksaeidi & Karami 2013; Jung et al. 2021).

In this review, we investigate the application of ML techniques in agricultural water management, with a focus on optimizing water allocation, enhancing decision-making, and mitigating climate-related risks. In the following section, we outline the key ML techniques employed in this domain and discuss their effectiveness based on recent studies.

Our objective is to synthesize key findings from these investigations, providing a comprehensive overview of the current state of knowledge in this rapidly evolving field. The novelty of this work lies in evaluating diverse ML techniques, emphasizing ensemble methods and real-time data integration, to enhance agricultural water management, sustainability, and resilience under climate variability.

ML offers several advantages for analyzing water usage patterns and optimizing resource allocation in agriculture (Sun & Scanlon 2019; Latif et al. 2023). First, it can handle large and complex datasets with high dimensionality, which are common in agricultural systems due to the multitude of interacting variables. Second, ML algorithms are capable of detecting subtle patterns and trends in data that may not be apparent through traditional statistical methods. This ability is particularly valuable for identifying factors influencing water usage and predicting future water demand. Additionally, ML models can adapt and improve over time as more data becomes available (Horvitz & Mulligan 2015), making them well suited for dynamic and evolving agricultural environments.

ML techniques have emerged as powerful tools in addressing the complexities of agricultural water usage. These techniques leverage algorithms and computational models to analyze vast datasets and extract valuable insights into water consumption patterns, irrigation practices, and crop water requirements (Elbeltagi et al. 2020b; Abdel-Hameed et al. 2024; Umutoni & Samadi 2024). By employing methods, such as predictive modeling, classification, and clustering, researchers, can develop accurate models that forecast future water demands, optimize irrigation scheduling, and identify areas for water conservation. Moreover, ML enables the integration of diverse data sources, including satellite imagery, weather forecasts, soil moisture measurements, and crop characteristics, to enhance the precision and efficiency of water management strategies in agriculture.

ML models have been widely used in water resource management due to their ability to process large datasets and extract meaningful insights. These models facilitate informed decision-making, optimize water allocation, and improv irrigation efficiency. Furthermore, ML enables the development of adaptive management frameworks that respond to changing environmental conditions and evolving agricultural needs. By integrating ML into water management systems, stakeholders can better understand the dynamics of water availability and usage, leading to more effective and resilient agricultural practices (Sun & Scanlon 2019; Fu et al. 2022).

Several ML algorithms are frequently employed in agricultural water management analysis. This discussion highlights the most commonly used ML models in the field, as reviewed in the literature, and their typical applications in agriculture. These models have been employed to address a wide range of challenges, from optimizing irrigation schedules to predicting crop yields and water footprints.

These constraints ensure that the predictions do not deviate from the actual values by more than the specified ⁠, allowing some flexibility through the slack variables for points that are otherwise hard to fit under this strict margin (Shrestha & Shukla 2015).

LightGBM is an efficient gradient boosting framework that uses tree-based learning algorithms. It is designed for distributed and efficient training, particularly on large datasets (Fan et al. 2019). LightGBM improves on traditional gradient boosting methods by using a histogram-based algorithm for faster processing and reduced memory usage. This model also handles large amounts of data with a higher accuracy and can efficiently manage categorical features as well (Ustuner & Sanli 2019; McCarty et al. 2020).

Regression trees (RT) are decision trees designed specifically for continuous outcome variables (regression problems). The tree splits the data into branches to form homogenous groups with similar target values. At each node, the tree chooses the split that minimizes the variance in the target variable, ultimately leading to a model that predicts the mean of each group. This method is particularly useful for capturing nonlinear relationships and interactions between features (Mokhtar et al. 2021).

Bagging, or bootstrap aggregating, is an EL technique that improves the stability and accuracy of ML models, particularly useful for high-variance models like decision trees (Schick et al. 2016; Winkler et al. 2018; Afrifa et al. 2022; Afrifa et al. 2023a). The process involves three main steps:

Bootstrap sample: Generate multiple subsets of the original dataset by sampling with replacement, creating diverse training sets (B1, B2, and B3).

Model training: Train individual models (M1, M2, and M3) on each subset in parallel.

This method reduces variance and helps prevent overfitting, enhancing the model's performance on unseen data. The diagram illustrates these steps, showing how data flows from the initial dataset through model training to the final aggregated output (Das et al. 2022).

Extreme Gradient Boosting (XGBoost) is a highly efficient, optimized library that improves upon traditional gradient boosting methods. It is notable for its ability to handle missing data, implement regularization to prevent overfitting, and prune trees. XGBoost operates by combining ‘weak’ learners into a ‘strong’ learner through an additive process, enhancing both speed and predictive accuracy, which is essential in various industry applications (Ge et al. 2022; Huber et al. 2022).

This streamlined approach makes XGBoost a preferred choice for achieving superior results in data-driven competitions and practical applications.

Various ML techniques, each with their own strengths, weaknesses, and typical applications in agriculture, are introduced in Table 1. This table provides a comprehensive summary derived from the review of several scholarly papers. The references for these studies, which discuss ML methods in detail, are elaborated upon in the subsequent section.

This section provides a detailed analysis of recent scholarly papers focusing on the application of ML techniques in agricultural water management. Each paper is meticulously reviewed to elucidate the methodologies employed, the specific agricultural contexts in which ML was implemented, and the resultant findings. The discussion aims to critically evaluate the effectiveness of these techniques in enhancing water management practices in agriculture. Additionally, the insights and conclusions drawn from these studies are synthesized to understand the broader implications and potential future directions for integrating ML in agricultural water management.

Geng et al. (2023) found that ML model performance in simulating crop water footprints varies by crop type, spatial scale, and water supply scenario. The RF model was most effective for maize, soybean, and rice at the site scale, while LightGBM performed better for wheat. At the provincial scale, LightGBM was the most accurate model. The study also shows higher accuracy of ML models in irrigated scenarios compared to rainfed ones, highlighting the importance of irrigation. Spatiotemporal features were the most significant variables, indicating their usefulness in rapid WF estimation when detailed soil data are unavailable.

Azzam et al. (2022) evaluated ML models for estimating green and blue water evapotranspiration (GWET and BWET) in the Amu Darya River Basin. The RF model outperformed others, achieving correlation coefficients up to 0.99 and root mean square error (RMSE) as low as 0.2637 mm/day. ANN and SVM models were also effective but less consistent. The inclusion of precipitation data significantly improved GWET predictions, emphasizing its importance.

Fan et al. (2021) examined the performance of various ML models for estimating daily maize transpiration in Northwest China. The models evaluated included SVM, XGBoost, ANNs, and deep neural networks (DNN). Using a combination of meteorological data (T_max, T_min, RH, U, Rs), soil water content (SWC), and leaf area index (LAI), the study found that the DNN model achieved the highest accuracy, with R² values ranging from 0.816 to 0.954 and RMSE values between 0.344 and 0.621 mm/day. This research underscores the effectiveness of DNNs in modeling complex nonlinear relationships in transpiration estimation, offering significant potential for enhancing irrigation practices and water management in agriculture.

Granata's (2019) findings reveal the robustness of ML models in predicting ET_a with high precision, where the M5P regression tree model exhibited the best performance across all tested models. This highlights the importance of a comprehensive input variable set in achieving high model accuracy, underscoring the critical role of detailed climatic and soil data in ET_a estimation. The study also demonstrates the variability in model performance based on the complexity of the input variables, with simpler models still performing satisfactorily but less accurately than more complex ones.

Shrestha & Shukla (2015) explored the use of ML models to estimate crop coefficient (K_c) and crop evapotranspiration (ET_c) for bell pepper and watermelon. The study compared the performance of SVM, ANNs, and relevance vector machines (RVM). The results demonstrated that the SVM model significantly outperformed the ANN and RVM models, achieving prediction errors of 2.6% for watermelon and 11.2% for bell pepper. The superior performance of the SVM model underscores its robustness and effectiveness in capturing the complex relationships between input variables and crop water use.

Elbeltagi et al. (2020b) evaluated various ML models for estimating the green and blue water footprints (GWFP and BWFP) of maize in the Nile Delta, Egypt. The study identified that ANNs achieved the highest predictive accuracy with coefficients of determination (R²) ranging from 0.94 to 0.99 and RMSE as low as 0.2637 mm/day. The inclusion of meteorological data such as temperature and precipitation significantly enhanced model performance, emphasizing the importance of comprehensive input data.

Mokhtar et al. (2021) investigated the performance of various ML models to estimate the blue and green water footprints (BWFP and GWFP) of rice in the Yunnan Province, southwest China. The study used four models: RT, RF, additive regression (AR), and reduced error pruning tree (REPT). Among these, the RT model in Scenario 1 (solar radiation, humidity, and vapor pressure deficit) outperformed others for BWFP estimation, with an RMSE value of 11.82 m³/ton and MAPE value of 0.5%. For GWFP, including precipitation in the input scenario significantly improved model accuracy. The study highlights the effectiveness of RT and RF models, particularly in scenarios with comprehensive climate and crop data, underscoring the potential of these models for reliable water footprint predictions in rice production under varying climatic conditions.

Geng et al. (2023) findings illustrate the potential of EL algorithms – RF, bagging, and adaptive boosting (Ad) – to estimate daily actual evapotranspiration (ET_a) in tea plantations using various meteorological data scenarios. The RF model demonstrated superior performance, with RMSE values ranging from 0.41 to 0.56 mm/day, MAE from 0.32 to 0.42 mm/day, and R² between 0.84 and 0.91. Bagging and SVM also showed good performance, though the MLP model was the least accurate. These results underscore the RF model's effectiveness and stability in predicting daily ETa, particularly with comprehensive climate and plant data inputs.

Ge et al. (2022) evaluated the performance of several ML models for predicting daily evapotranspiration (ET) in greenhouse-grown tomatoes. The XGBoost model achieved the highest prediction accuracy, with a mean square error (MSE) of 0.032, RMSE of 0.163, and a coefficient of determination (R²) of 0.981. GBR and SVR also showed good performance but did not match the accuracy of XGBoost. This study underscores the importance of comprehensive input data, including net solar radiation, temperature, and relative humidity, to enhance model performance. Ge et al. highlight the potential of these ML models, particularly XGBoost, in optimizing irrigation practices and improving water resource management in greenhouse agriculture.

Elbeltagi et al. (2020a) demonstrate that DNNs effectively predict the water footprint under varying climatic scenarios, emphasizing the importance of precise and adaptable modeling techniques in agriculture. The study illustrates that DNN models can capture the complex interactions between climate factors and crop water use, providing reliable predictions that are crucial for strategic agricultural planning. The findings suggest that while DNNs offer significant advantages in modeling water footprints, their effectiveness depends on the quality of input data and the specific configurations of the neural network. This research underscores the potential of ML to enhance predictive accuracy and aid in the development of resilient agricultural practices against the backdrop of global climate change.

Dehghanisanij et al. (2022) demonstrated that a hybrid ML approach combining an adaptive neuro-fuzzy inference system (ANFIS) and seasonal optimization (SO) algorithm effectively estimates water-use efficiency (WUE) and yield in apple orchards. The SO–ANFIS model outperformed traditional methods with R² values of 0.989 for WUE and 0.988 for yield. The study highlighted the model's ability to handle complex irrigation and climatic data, significantly enhancing prediction accuracy and supporting efficient water management practices.

Virnodkar et al. (2020) emphasize the superior capabilities of ML models, such as ANN, SVM, and RF, over traditional RS methods in detecting crop water stress with higher accuracy and efficiency. They explored the integration of remote sensing and ML techniques to determine crop water stress. The study demonstrated that ML models, particularly SVM and ANNs, significantly enhance the accuracy of water stress predictions compared to traditional methods. These models effectively process large volumes of remote sensing data, offering precise and efficient crop water stress assessments. The findings highlight the potential of combining remote sensing with ML to improve water management and irrigation practices in precision agriculture.

Gao et al. (2023) provided a comprehensive review of the applications of MLin irrigation, emphasizing its potential to enhance agricultural water productivity. The study categorized ML applications into water scarcity diagnosis, water demand prediction, and irrigation decision-making, showcasing the effectiveness of models such as CNN, RF, LSTM, SVM, and XGBoost. These models demonstrated high accuracy in diagnosing water stress, predicting soil moisture content, and optimizing irrigation schedules, thereby significantly improving water use efficiency and crop yields. Despite the promising results, Gao et al. highlighted challenges such as data quality, model interpretability, and portability. To address these issues, they proposed a unified ML framework integrating deep learning and optimization algorithms to enhance model robustness and applicability across various agricultural contexts.

Liakos et al. (2018) demonstrate that ML technologies, when combined with the appropriate sensors and data, can significantly improve water management in agriculture. They highlighted that ML models, such as ANN, ELM, and SVM, significantly improved the estimation of evapotranspiration and soil moisture content, achieving high accuracy and efficiency. For example, regression models achieved a correlation coefficient (R) of 0.9999 in estimating monthly mean evapotranspiration in arid regions. Despite these successes, challenges like data quality and model interpretability remain. Liakos et al. emphasized the need for integrating ML with big data and high-performance computing to address these issues, underscoring ML's transformative potential in enhancing precision agriculture and sustainable water management.

Benos et al. (2021) conducted a comprehensive review of ML applications in agriculture, focusing on optimizing irrigation and enhancing water use efficiency. The study highlighted that ML models, such as ANNs, SVMs, and CNNs, significantly improved the accuracy of predicting evapotranspiration, soil moisture, and irrigation scheduling. The review underscored ML's potential to advance precision agriculture and sustainable water management, despite challenges like data quality and model interpretability.

Umutoni & Samadi (2024) illustrate the significant advancements ML has brought to irrigation practices, notably through predictive models that utilize data from diverse sources such as soil sensors and weather forecasts. The paper evaluates the strengths and limitations of various ML algorithms in real-world applications, noting that while ML significantly enhances decision-making, its effectiveness hinges on data quality, algorithm suitability, and the integration of expert knowledge. The discussion underscores the need for ongoing research to address data scarcity, improve model accuracy, and ensure the practical deployment of ML systems in diverse agricultural settings.

Abdel-Hameed et al. (2024) investigated the estimation of potato water footprint in arid regions using ML models. The research shows that models such as RF, ANN, and SVM can significantly enhance the accuracy of water footprint predictions by incorporating climatic, soil, and spatiotemporal data. The findings indicate that the RF model is particularly effective for site-scale predictions, while ANN models are robust across various climatic conditions. The study underscores the importance of detailed input data and site-specific modeling to improve the precision of water footprint estimates.

Elbeltagi et al. (2023) discovered that the performance of ML models in predicting vapor pressure deficit (VPD) varies significantly across different semi-arid regions in Egypt, which are crucial for optimizing water use in agriculture. Utilizing several ML models, including Linear Regression, Random SubSpace, REPTree, and notably RF, the study rigorously tested these models to determine their effectiveness in forecasting VPD. The conclusion drawn was that the RF model, due to its high accuracy and robust performance across complex datasets, was the most effective tool for this purpose.

Kumar et al. (2020) explore the use of ANN models, specifically K-SOM and FF-BP, to predict the crop water stress index (CWSI) for Indian mustard. The study demonstrates the high accuracy of these models in estimating CWSI, with K-SOM achieving R² values of 0.97 during development and 0.96 during validation phases, indicating low bias error. In comparison, the FF-BP model showed R² values of 0.861 and 0.747 for the development and validation phases, respectively, with a higher bias error. The research underscores the importance of using precise input data, including air and canopy temperatures and relative humidity, to enhance model performance. Kumar et al. highlight the potential of these ANN models in improving water resource management by providing accurate stress index predictions, thus aiding in optimizing irrigation practices.

Das et al. (2022) evaluated the effectiveness of various ML models for surface soil moisture mapping using combined optical, thermal, and microwave remote sensing data. Focusing on crops such as wheat, maize, mustard, and pulses in a New Delhi farm, the study compared Cubist, RF, gradient boosting machine (GBM), and a stacking ensemble approach. The stacking model, integrating outputs from Cubist, RF, and GBM with the elastic net as a meta-learner, achieved the lowest mean bias error (MBE) of 0.18% and RMSE value of 5.03%. The RF model stood out with the highest correlation (r = 0.71) and lowest RMSE (5.17%), underscoring its superior predictive performance. This research underscores the potential of advanced ML models, particularly ensemble techniques, in enhancing soil moisture estimation accuracy, essential for optimized irrigation and water management.

In this section, Table 2 presents a rigorous comparative analysis of various ML models applied in the field of agricultural water management. This table meticulously details each model's performance metrics, including accuracy, computational efficiency, scalability, and adaptability to diverse agricultural conditions. By examining these critical parameters, the table offers an in-depth understanding of the strengths and limitations of each model. This comprehensive summary facilitates the identification and selection of the most suitable ML approaches for addressing specific water management challenges in agriculture, thereby contributing to the optimization of water usage and sustainability in agricultural practices.

ML has emerged as a game-changer in agricultural water management, enhancing irrigation efficiency, water demand forecasting, and climate resilience. This study underscores the advantages of ensemble methods like RF and LightGBM for handling large datasets and the strength of ANNs in specific water-related predictions. However, challenges such as data availability, model interpretability, and scalability must be addressed to fully harness ML's potential. Future efforts should focus on integrating advanced sensor networks, refining AI-driven decision-making tools, and ensuring accessibility for diverse agricultural stakeholders. By overcoming these barriers, ML-driven solutions can play a pivotal role in sustainable and resilient water management practices worldwide.

• 1. Data quality and availability: While ML provides numerous advantages for agricultural water management, several challenges must be addressed to maximize its potential. A detailed examination of these limitations can guide effective solutions. Key areas where ML might face limitations include data quality, model interpretability, scalability, and climate variability (Figure 7) (L'Heureux et al. 2017; Zhou et al. 2017; Weersink et al. 2018; Cravero et al. 2022; Gao et al. 2023).

Challenge: ML algorithms require large volumes of high-quality data to function optimally. In many agricultural settings, especially in developing countries, data may be sparse, inaccurate, or outdated. This can lead to poor model performance and unreliable predictions.

Potential solution: Initiatives to improve data collection infrastructure, such as government-funded projects to deploy more sensors and satellites, can enhance data quality. Collaborative efforts between academic institutions, industry, and government to create open-source data repositories could also be beneficial (Kamilaris & Prenafeta-Boldú 2018).

• 2. Model complexity and interpretability:

Challenge: Many advanced ML models, such as DNNs, are often seen as ‘black boxes’ due to their complex nature, making it difficult for practitioners to trust and interpret their predictions.

Potential solution: Research into explainable AI (XAI) is vital, as these techniques can make the workings of complex models more transparent and understandable to non-experts. Developing hybrid models that combine ML with traditional decision-making frameworks could also help enhance trust and interpretability (Ryo 2022).

• 3. Integration with existing systems:

Challenge: Integrating ML systems with existing agricultural management practices can be challenging due to compatibility issues, the need for technical expertise, and resistance from traditional practitioners.

Potential solution: Developing user-friendly ML tools that require minimal setup and are compatible with existing technologies can ease integration. Training sessions and workshops for farmers and agricultural managers can help bridge the knowledge gap and promote acceptance (Wolfert et al. 2017).

• 4. Scalability and adaptability:

Challenge: ML models developed in one region or for one type of crop may not perform well when applied to different regions or crops due to varying climatic and soil conditions.

Potential solution: Focused research on developing adaptable and scalable ML models that can adjust to different environmental conditions and agricultural practices is essential. Using transfer learning and fine-tuning models based on local data can improve scalability (Liakos et al. 2018).

• 5. Cost and resource requirements:

Challenge: The deployment of sophisticated ML models can be resource-intensive, requiring significant computational power and expertise, which might be unaffordable for small to medium-sized farms.

Potential solution: Cloud-based ML solutions can reduce the need for on-site computational resources. Subsidies or financial incentives from governments or agricultural cooperatives can help smaller farms adopt these technologies (Shamshiri et al. 2018).

• 6. Regulatory and ethical concerns:

Challenge: The use of ML in agriculture raises concerns about data privacy, ownership, and ethical use, especially when data from multiple sources are integrated and analyzed.

Potential solution: Establishing clear regulations and ethical guidelines for the use of agricultural data and ML models can help address these concerns. Stakeholder engagement in developing these regulations can ensure that they are broadly acceptable and effective (Ryan 2023).

• 7. Climate change and uncertainty in future projections

Challenge: Climate variability introduces significant uncertainties in ML-based water management models. Changes in temperature, precipitation patterns, and extreme weather events can disrupt model assumptions, reducing their predictive reliability.

Potential solution: Hybrid climate-aware ML models that integrate hydrological simulations with AI-driven predictions can enhance robustness against climate uncertainties. Additionally, ensemble modeling approaches that aggregate predictions from multiple ML models can improve reliability in uncertain scenarios (Azmat et al. 2022).

ML has increasingly become a pivotal technology in revolutionizing agricultural practices, particularly in water management. This review has illustrated that ML offers substantial promise for enhancing the productivity and sustainability of agricultural operations through precise, efficient water management – crucial in the face of mounting global challenges such as water scarcity and climate change.

Our extensive examination of ML applications in agricultural water management showcases these technologies' transformative capabilities. By integrating a variety of models such as RF, ANNs, SVM, and LightGBM, significant strides have been made in improving prediction accuracy and management strategies across diverse agricultural environments. These models facilitate nuanced decision-making, optimize resource allocation, and enable robust risk mitigation under varying climatic conditions.

Key findings from the papers reviewed reveal:

• Predictive accuracy: ML models, particularly ensemble methods such as RF and LightGBM, deliver exceptional predictive accuracy, crucial for effective irrigation planning and water allocation.

• Data integration and analysis: The successful integration of various data types, including remote sensing and ground sensor data, enhances model reliability and broadens their applicability, supporting comprehensive water management strategies.

• Model-specific insights: RF and ANN excel in handling complex datasets and generating dependable predictions for water consumption and evapotranspiration – essential for informed decision-making in regions facing water scarcity. LightGBM's effectiveness underscores the importance of selecting models based on specific agricultural needs and environmental conditions.

• Despite these advances, several challenges persist: High-quality, diverse datasets are essential for training robust models, and the complexity of integrating ML technologies into existing agricultural frameworks remains a significant hurdle. Additionally, ensuring model interpretability and managing socio-economic factors are critical for the practical application of these solutions.

To propel this field forward, future research should focus on:

• - Enhancing data collection techniques: Improved data quality and granularity will aid in refining ML models, making them more responsive to specific agricultural contexts.

• - Improving model transparency and usability: Developing user-friendly models with clear methodologies will facilitate broader adoption, especially by practitioners who are not ML experts.

• - Tailoring models to specific needs: Customizing ML approaches to accommodate particular crop types, climates, and water management systems will boost their effectiveness and relevance.

• - Fostering interdisciplinary collaborations: Engaging various scientific and industry disciplines can spur innovations that blend agronomic knowledge with cutting-edge ML techniques, leading to integrated solutions for water management.

In conclusion, ML not only offers a promising avenue for advancing agricultural water management practices but also plays a critical role in shaping the future of global agriculture. By continuing to leverage and refine these technologies, there is tremendous potential to improve the sustainability and efficiency of farming operations worldwide, ensuring food security and resource conservation in an increasingly unpredictable global climate. This synthesis provides a clear overview of the current capabilities of ML in agricultural water management and sets a comprehensive roadmap for ongoing research and practical applications.

All authors have reviewed and approved the final manuscript and consent to its publication.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

M. M. and F. M. wrote the original draft, conducting the literature review, and designing figures, tables, and illustrations. D. B. synthesizing information from various sources, enhancing readability and flow, offering suggestions for improvement, overseeing final revisions, supervising the research process, and ensuring methodological rigor. S. M., A. D., and J. L. N. reviewing and revising sections related to their expertise, providing critical feedback and suggestions, ensuring technical accuracy, and contributing to the interpretation of findings. S. M. and H. G. reviewed and edited the article, revision of the manuscript in response to reviewers' comments.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.