Securing freshwater for all

"Distorted Lake Trees" by Lone Thomasky & Bits&Bäume via betterimagesofai.org. (CC-BY 4.0).

Fresh water, the bloodstream of the biosphere, is fundamental for all life on land and all societal functions. Can AI be used to map, analyze, and protect freshwater flows on a changing planet?

Introduction

Freshwater—“the bloodstream of the biosphere”— is fundamental for all life on land and all societal functions. The availability and dynamics of freshwater play a critical role in the regulation of Earth’s climate system and, at the same time, are among the most visible manifestations of climate change. Droughts and extreme precipitation events are becoming more severe, intense, prolonged, and abrupt, resulting in costly damages to agriculture, infrastructure, and livelihoods, with implications for migration patterns, conflicts, and security. Shifting hydroclimate predisposes ecosystems to regime shifts, thereby increasing the risk of disease outbreaks. Four billion people are estimated to face severe water scarcity conditions at least one month per year.1 Freshwater flows in the landscape and across the atmosphere do not respect administrative borders, complicating decision-making.

Hence, securing freshwater for all demands a comprehensive understanding of complex system dynamics spanning diverse scales and sectors, which necessitate high-quality data and modeling, as well as just and effective decision-making processes. Conventional approaches to supporting freshwater sustainability are hampered by data scarcity in many regions, slow model development, and the inherent complexity of decision processes.2,3 Artificial intelligence (AI) can accelerate progress by bridging data limitations, enhancing model accuracy and speed, and facilitating informed decision-making. While AI can play a transformative role in strengthening freshwater’s role for resilience and sustainability, proactive consideration of emerging ethical and governance challenges that arise with its application is crucial for ensuring equitable outcomes.

In this chapter, we illustrate some applications of AI in agriculture, including disaster reduction, water resources management for forecasting, understanding water cycle dynamics, and supporting flood and drought prediction. In all these areas, the application of AI has demonstrated usefulness and continues to develop rapidly. The ethical and governance challenges that arise from the use of AI in freshwater management will also be briefly discussed.

Several key aspects of water sustainability, beyond the scope of this chapter, include AI applications for ensuring access to clean water and sanitation, addressing water quality and aquatic ecosystem health, water in industries and energy production, and understanding water-induced collapse and tipping points.

AI for Freshwater Data

Informed decisions and management of water resources require accurate water-relevant data at the proper resolution. As several studies have demonstrated, AI can have a transformative impact on water sustainability in multiple ways, including:

Counteracting geographical bias. Ground-based data collection of freshwater-related variables is unevenly distributed across the world, which has detrimental implications for security in some of the world’s most socioeconomically vulnerable regions. Many regions in sub-Saharan Africa, parts of Asia, and South America lack comprehensive monitoring networks for basic hydrological and hydroclimatic data, including precipitation, evaporation, and runoff. For example, a lack of warning systems and preparedness contributed to making flood events in Africa four times more deadly than in Europe or the US in recent years (2020–2022).4 AI methods, particularly supervised and semi-supervised machine learning (ML) approaches, offer diverse opportunities to contribute to water-related data generation and collection, and can help mitigate global sampling biases.5 Because differences in data formats and standards can make it challenging to combine datasets from different regions, AIbased approaches can complement traditional statistical techniques (e.g., kriging, a method for spatial interpolation) to improve data integration.6 Verification methods that combine satellite observations with local ground measurements can also enhance data quality for AI-based hydrological models, especially in regions with limited data coverage.7,8

Increasing the quality and speed of data analyses. A lack of freshwater-related data (i.e., accurate, continuous, and with high resolution) hinders the monitoring, assessment, and prevention of unsustainable water use and other adverse human impacts. Reanalysis datasets of hydroclimatic variables combine observations with model outputs to create consistent, globally gridded time-series data. They are useful for regional and global applications but often lack sufficient granularity for local applications.9 Remote sensing provides high-resolution observational data but remains limited by uncertainty. This is especially true for variables such as evaporation, which are often indirectly derived and rely on further assumptions. Recent efforts have shown promise in addressing these limitations with AI techniques. For example, deep neural networks have been used to improve the accuracy and reliability of precipitation data from reanalysis products,10 and to map and classify surface water bodies such as lakes, wetlands, and reservoirs from remote sensing data.11,12 Around a third of remote sensing-based studies on surface water detection and delineation used AI.13 However, many hydrologically relevant variables, such as soil moisture and evapotranspiration, are not readily measurable from optical satellite data and require calibration with ground measurements.5 Combining networks of ground measurement stations with remote sensing data and supervised ML techniques allows for the generation and extension of high-resolution, spatially and temporally continuous datasets for these variables.14 Further, AI can also be essential in the timely delivery of information for emergency responses, for example through fully automatic detection of flood extents from satellite imagery.15,16

Enabling the collection of new types of data. Social media and citizen science measurement networks provide additional, novel data sources, particularly for real-time assessments and early warnings of risks such as floods.17,18 Some recent research has focused on developing AI-based early warnings of flood impacts using text posts from social networks,19–22 but given the challenges in accessibility of mostly privately held data and often a lack of georeferencing, such applications have not been widely employed yet. Supervised ML models such as deep neural networks in combination with remote sensing data also allow for detecting and mapping hydrologically relevant data that are otherwise not easily established, such as different irrigation schemes on agricultural lands.23

AI-Assisted Predictive Modeling

Modeling is crucial for securing freshwater sustainability. Models (i.e., conceptual models, mechanistic models, process-based general circulation models) can help test hypotheses and support a better understanding of hydrological dynamics and processes of varying complexity. Operational models can, as an additional example, be used for forecasting and early warning of water-related disasters (such as drought and flood prediction). AI can help to improve predictions and understanding gained from freshwater modeling efforts by enhancing model components (e.g., parameterization, equations) or simulation outputs (post-processing of model outputs), allowing for alternative modes of modeling that are computationally less expensive and/or more accurate (e.g., data-driven, surrogate, and hybrid models). Some of the important contributions of predictive modeling for securing freshwater provisioning are presented in more detail here:

Enhancing model parameters, equations, and outputs. Key applications of AI in this context include identifying optimal parameter values for specific subprocesses in process-based models and developing context-specific bias correction algorithms for predicting hydroclimatic variables.24 For advancing basic scientific inquiry, deep learning (DL) may be leveraged to enable data-driven equation discovery.25 AI approaches have been utilized for the spatiotemporal downscaling of coarse outputs from process-based models,26 for example using neural networks to increase the spatial resolution of precipitation predictions by a factor of up to 25.27 This generates higher-resolution and regionally accurate predictions of the key climatic and hydrological processes, which are necessary for local management plans.

Diversifying alternative modeling approaches. In data-driven modeling, AI is used to directly derive relationships based on empirical data, thereby better accounting for complexity not captured in models,28 or to enhance model ensemble data when there are limited observations, thereby saving computational time.29 AI-based surrogate models are trained on the input and output parameters of process-based models, learning to provide faster and less computationally expensive predictions of expected climatic or hydrological conditions under specific scenarios. They are particularly useful for assessing model sensitivity, as well as the risks associated with a wide range of possible scenarios.30 In hybrid modeling approaches, AI algorithms have been integrated and combined with process-based models in various ways to enhance their performance, interpretability, and applicability for decision-making. One common strategy is to use AI-based post-processing of hydrological model output to better capture local ground truth conditions. Generally, the process-based submodels are useful for well-understood physical processes, while AI methods are employed for fine-tuning and representing complex and poorly described subprocesses.31 Such approaches have proven valuable in enhancing flood prediction,32 groundwater simulations,33 and drought forecasting.34

AI-Assisted Decision-Making

AI can be used throughout the entire decision-making chain, from data retrieval and sense-making to facilitating discussions on risks.35 High speed and accuracy in forecasting are critical for saving lives in the context of flood and drought disasters. AI methods, such as the coupling of artificial neural networks, the Bayesian framework, and genetic algorithms, can also directly assist decision-making for water sustainability by improving the use and allocation of water resources, forecasting the consumption and demand of water resources,36 facilitating communication and dialogues, and enhancing the understanding of risks and uncertainty. Among others, AI approaches such as ML, artificial neuro-genetic networks, and AI-powered drones and satellites have been used to forecast, monitor, and optimize irrigation in agriculture, reservoirs and dams management, and industrial and household water use.37–39

Moreover, communication and dialogue are essential in facilitating decision-making in multi-stakeholder contexts, which water decisions inevitably are. Large language models (LLMs) can, for example, be used as moderators in meeting contexts to help navigate complex issues and/or provide information that can be perceived more neutrally in contexts with diverse views and needs, such as urban planning.40 Furthermore, AI can also aid in understanding risks and uncertainty by simulating multiple scenarios, analyzing larger sets of data, and generating visualizations. For example, WaterGPT is an LLM designed to act as a hydrology expert by processing and analyzing images and text, and answering questions in natural language.41 General-purpose multimodal LLMs such as GPT-4 Vision, Gemini, etc., showed potential in various hydrological applications.42

Limitations and Key Challenges

The full transformative potential of AI for securing freshwater for both people and the planet remains to be explored, and key limitations and challenges need to be addressed. Critically, AI needs to be more understandable by users, and its continued application should pay close attention to ethical implications, continued investments in ground measurements, and the considerable water requirements of the data centers used to power AI. Some of the more pressing concerns include:

Explainability and transparency. AI models, especially DL algorithms, are highly complex and opaque. DL models often act as “black boxes,” making it difficult to understand the causal mechanisms behind the results and build trust among decision-makers, especially since the academic experts developing these models may not be able to explain outcomes. Understanding, transparency, and trust are particularly relevant in water management, where decisions impact resource allocation, flood prevention, and climate response strategies. Developing models with greater transparency, such as decision trees43 or SHapley Additive exPlanations (SHAP),44 can provide clear insights that help water managers understand key factors driving hydrological variability.

Access, fairness, and other ethical concerns. The computational demands of AI models also present a challenge. Training and optimizing complex models requires substantial computing power, particularly for large-scale simulations and high-resolution predictions.44 This can pose challenges for the use and development of AI in resource-limited settings. Ensuring that AI technologies are accessible to all regions, developed in collaboration with local agencies, and tailored to local use cases, particularly in low-income countries, is crucial to prevent widening disparities in water resource management.45,46 Stakeholders and policymakers also face challenges when adopting AI tools in water research. A lack of technical expertise can hinder understanding and adoption of AI models, particularly for non-specialist decision-makers. Furthermore, AI models may produce results that conflict with traditional knowledge systems, creating skepticism and reducing trust. Ensuring that AI tools follow principles of trustworthy AI,47 and are accompanied by clear documentation on how they can be used, simple user interfaces, and precise documentation of trained resources can improve stakeholder engagement and confidence.48,49

Fundamental lack of observations. Fundamentally, AI relies on data inputs, and a key challenge is the lack of observations. Existing biases in data collection are particularly problematic for addressing water issues, as water processes and dynamics are highly heterogeneous and context-dependent.38 A key challenge lies in the availability of evenly distributed, high-quality, and unbiased data. For most regions, hydrological data are often incomplete in both space and time, can be inconsistent in their measurement, and are lacking in regions with limited monitoring capabilities. Such gaps in data reduce the development and reliability of AI-based models.50

Negative impacts of AI usage on water sustainability. While AI can be used for good, the explosive growth in AI use by a variety of users has also necessitated a significant increase in data centers that require water for cooling. Training the GPT-3 language model has been estimated to lead to the evaporation of 700,000 liters of clean freshwater, and globally, AI is projected to lead to 4.2–6.6 billion m3 of water withdrawal by 2027.51 Such estimates are highly uncertain due to a lack of transparency from the tech companies that train AI, and this uncertainty can hopefully be lessened with improvements in computationally efficient algorithms, increased public awareness, and legislation.52

Addressing these challenges will require collaboration across various disciplines, from academic developers to applying stakeholders, as well as improved openness to data sharing from both sides. This will also involve the development of transparent AI models that align with the complexities of hydrological systems.49

Bibliography

  1. Mekonnen, M. M. & Hoekstra, A. Y. Four billion people facing severe water scarcity. Adv. 2, e1500323 (2016).
  2. Kratzert, F. et al. Toward Improved Predictions in Ungauged Basins: Exploiting the Power of Machine Learning. Water Resour. Res. 55, 11344–11354 (2019).
  3. Smits, A. P. et al. Too much and not enough data: Challenges and solutions for generating information in freshwater research and monitoring. Ecosphere 16, e70205 (2025).
  4. Tzachor, A. et al. How to reduce Africa’s undue exposure to climate risks. Nature 620, 488–491 (2023).
  5. Xu, T. & Liang, F. Machine learning for hydrologic sciences: An introductory overview. WIREs Water 8, e1533 (2021).
  6. Zhou, L. Perspective Chapter: Big Data and Deep Learning in Hydrological Modeling. in Earth Sciences (ed. Guo, C.) vol. 4 (IntechOpen, 2024).
  7. Jiang, P., Shuai, P., Sun, A., Mudunuru, M. K. & Chen, X. Knowledge-informed deep learning for hydrological model calibration: an application to Coal Creek Watershed in Colorado. Earth Syst. Sci. 27, 2621–2644 (2023).
  8. Bhowmik, A., Moulik, A. & Majhi, T. Comparative study of artificial neural network and hydrological models towards runoff estimation. J. Adv. Biochem. Res. 8, 108–112 (2024).
  9. Seyyedi, H., Anagnostou, E. N., Beighley, E. & McCollum, J. Satellite-driven downscaling of global reanalysis precipitation products for hydrological applications. Earth Syst. Sci. 18, 5077–5091 (2014).
  10. Cavaiola, M., Tuju, P. E. & Mazzino, A. Accurate and efficient AI-assisted paradigm for adding granularity to ERA5 precipitation reanalysis. Rep. 14, 26158 (2024).
  11. Pekel, J.-F., Cottam, A., Gorelick, N. & Belward, A. S. High-resolution mapping of global surface water and its long-term changes. Nature 540, 418–422 (2016).
  12. Hao, Z. et al. GRDL: A New Global Reservoir Area‐Storage‐Depth Data Set Derived Through Deep Learning‐Based Bathymetry Reconstruction. Water Resour. Res. 60, e2023WR035781 (2024).
  13. Bijeesh, T. V. & Narasimhamurthy, K. N. Surface water detection and delineation using remote sensing images: a review of methods and algorithms. Water Resour. Manag. 6, 68 (2020).
  14. Han, Q. et al. Global long term daily 1 km surface soil moisture dataset with physics informed machine learning. Data 10, 101 (2023).
  15. Nemni, E., Bullock, J., Belabbes, S. & Bromley, L. Fully Convolutional Neural Network for Rapid Flood Segmentation in Synthetic Aperture Radar Imagery. Remote Sens. 12, 2532 (2020).
  16. Bentivoglio, R., Isufi, E., Jonkman, S. N. & Taormina, R. Deep learning methods for flood mapping: a review of existing applications and future research directions. Earth Syst. Sci. 26, 4345–4378 (2022).
  17. Aggarwal, C. C. & Abdelzaher, T. Social Sensing. in Managing and Mining Sensor Data (ed. Aggarwal, C. C.) 237–297 (Springer US, Boston, MA, 2013). doi:10.1007/978-1-46146309-2_9.
  18. Arthur, R., Boulton, C. A., Shotton, H. & Williams, H. T. P. Social sensing of floods in the UK. PLOS ONE 13, e0189327 (2018).
  19. Smith, L., Liang, Q., James, P. & Lin, W. Assessing the utility of social media as a data source for flood risk management using a real‐time modelling framework. Flood Risk Manag. 10, 370–380 (2017).
  20. Restrepo-Estrada, C. et al. Geo-social media as a proxy for hydrometeorological data for streamflow estimation and to improve flood monitoring. Geosci. 111, 148–158 (2018).
  21. Puttinaovarat, S. & Horkaew, P. Flood Forecasting System Based on Integrated Big and Crowdsource Data by Using Machine Learning Techniques. IEEE Access 8, 5885–5905 (2020).
  22. Costa, J. P. et al. Integrating Social Media, News and Machine Learning for Enhanced Hydrological Event Detection and Management. in 15th International Conference on Hydroinformatics 314–315 (The International Association for Hydro-Environment Engineering and Research (IAHR), 2024). doi:10.3850/iahr-hic2483430201-278.
  23. Paolini, G., Escorihuela, M. J., Merlin, O., Sans, M. P. & Bellvert, J. Classification of Different Irrigation Systems at Field Scale Using Time-Series of Remote Sensing Data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 15, 10055– 10072 (2022).
  24. Sun, A. Y. & Scanlon, B. R. How can Big Data and machine learning benefit environment and water management: a survey of methods, applications, and future directions. Res. Lett. 14, 073001 (2019).
  25. Song, W. et al. Towards data-driven discovery of governing equations in geosciences. Earth Environ. 5, 589 (2024).
  26. Wang, F., Tian, D., Lowe, L., Kalin, L. & Lehrter, J. Deep Learning for Daily Precipitation and Temperature Downscaling. Water Resour. Res. 57, e2020WR029308 (2021).
  27. Wang, F. & Tian, D. On deep learning-based bias correction and downscaling of multiple climate models simulations. Dyn. 59, 3451–3468 (2022).
  28. Jones, A. et al. AI for climate impacts: applications in flood risk. Npj Clim. Atmospheric Sci. 6, 63 (2023).
  29. Gibson, P. B. et al. Training machine learning models on climate model output yields skillful interpretable seasonal precipitation forecasts. Earth Environ. 2, 159 (2021).
  30. Wagenaar, D. et al. Invited perspectives: How machine learning will change flood risk and impact assessment. Hazards Earth Syst. Sci. 20, 1149–1161 (2020).
  31. Kraft, B., Jung, M., Körner, M., Koirala, S. & Reichstein, M.Towards hybrid modeling of the global hydrological cycle. Hydrol. Earth Syst. Sci. 26, 1579–1614 (2022).
  32. Nearing, G. et al. AI Increases Global Access to Reliable Flood Forecasts. Preprint at https://doi.org/10.48550/ARXIV.2307.16104 (2023).
  33. Moazenzadeh, R. & Izady, A. A hybrid calibration method for improving hydrological systems using ground-based and remotely-sensed observations. Hydrol. 615, 128688 (2022).
  34. Mishra, A. K., Desai, V. R. & Singh, V. P. Drought Forecasting Using a Hybrid Stochastic and Neural Network Model. Hydrol. Eng. 12, 626–638 (2007).
  35. Reichstein, M. et al. Early warning of complex climate risk with integrated artificial intelligence. Commun. 16, 2564 (2025).
  36. González Perea, R., Camacho Poyato, E., Montesinos, P. & Rodríguez Díaz, J. A. Optimisation of water demand forecasting by artificial intelligence with short data sets. Eng. 177, 59–66 (2019).
  37. Perea, R. G., Poyato, E. C., Montesinos, P. & Díaz, J. A. R. Irrigation Demand Forecasting Using Artificial Neuro-Genetic Networks. Water Resour. Manag. 29, 5551–5567 (2015).
  38. Ashoka, P. et al. Artificial Intelligence in Water Management for Sustainable Farming: A Review. Sci. Res. Rep. 30, 511–525 (2024).
  39. Jayakumar, D., Bouhoula, A. & Al-Zubari, W. K. Unlocking the Potential of Artificial Intelligence for Sustainable Water Management Focusing Operational Applications. Water 16, 3328 (2024).
  40. Zhou, Z., Lin, Y., Jin, D. & Li, Y. Large Language Model for Participatory Urban Planning. Preprint at https://doi. org/10.48550/ARXIV.2402.17161 (2024).
  41. Ren, Y. et al. WaterGPT: Training a Large Language Model to Become a Hydrology Expert. Water 16, 3075 (2024).
  42. Kadiyala, L. A., Mermer, O., Samuel, D. J., Sermet, Y. & Demir, I. The Implementation of Multimodal Large Language Models for Hydrological Applications: A Comparative Study of GPT-4 Vision, Gemini, LLaVa, and Multimodal-GPT. Hydrology 11, 148 (2024).
  43. Blockeel, H., Devos, L., Frénay, B., Nanfack, G. & Nijssen, S. Decision trees: from efficient prediction to responsible AI. Artif. Intell. 6, 1124553 (2023).
  44. Huerta, E. A. et al. Convergence of artificial intelligence and high performance computing on NSF-supported cyberinfrastructure. Big Data 7, 88 (2020).
  45. Heeks, R. Digital inequality beyond the digital divide: conceptualizing adverse digital incorporation in the global South. Technol. Dev. 28, 688–704 (2022).
  46. Lehdonvirta, V., Wú, B. & Hawkins, Z. Compute North vs. Compute South: The Uneven Possibilities of Compute-based AI Governance Around the Globe. AAAIACM Conf. AI Ethics Soc. 7, 828–838 (2024).
  47. Ethics Guidelines for Trustworthy AI. (2019).
  48. Kaur, D., Uslu, S., Rittichier, K. J. & Durresi, A. Trustworthy Artificial Intelligence: A Review. ACM Comput. Surv. 55, 1–38 (2023).
  49. Kinney, M., Anastasiadou, M., Naranjo-Zolotov, M. & Santos, V. Expectation management in AI: A framework for understanding stakeholder trust and acceptance of artificial intelligence systems. Heliyon 10, e28562 (2024).
  50. Tang, G., Wood, A. W. & Swenson, S. On AI-based large-sample emulators for land/hydrology model calibration and regionalization. Preprint at https://doi.org/10.22541/essoar.173386122.27239661/v1 (2024).
  51. Li, P., Yang, J., Islam, M. A. & Ren, S. Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models. Preprint at https://doi.org/10.48550/ ARXIV.2304.03271 (2023).
  52. Garcia, M. AI Uses How Much Water? Navigating Regulation Of AI Data Centers’ Water Footprint Post-Watershed Loper Bright Decision. Preprint at https://doi.org/10.2139/ ssrn.5064473 (2025).
Topics: Leveraging technologies
Published: 2025-11-05

News & events

See more news