Building on quicksand: the paradox of AI advances and declining in situ networks in hydrometeorology

Over the past few years, the application of artificial intelligence (AI) in hydrometeorology has expanded rapidly, encompassing tasks such as the prediction of weather extremes and streamflow, uncertainty quantification, and data fusion to enhance data quality and spatial or temporal resolution^1,2,3. To date, much of this progress has been concentrated in high-income countries. However, there is increasing anticipation in low- and middle-income countries (LMICs) that AI may serve as a catalyst for development through technological “leapfrogging”^4,5. This concept posits that less developed economies may adopt emerging technologies more readily, thereby bypassing intermediate stages of technological development⁶.

The effectiveness of AI arises from its ability to learn complex yet systematic relationships from training data, enabling the construction of models that map high-dimensional inputs to outputs. As a result, the quality and representativeness of the training datasets are critical determinants of model performance. Errors, uncertainties, or biases present in the data may be incorporated into, and potentially amplified by, AI models. Such amplification can manifest as so-called “butterfly effects”, whereby small biases or distributional shifts in the input data propagate through nonlinear model architectures, high-dimensional feature spaces, and feedback mechanisms, leading to disproportionately large and sometimes unpredictable outcomes.

Predicting and providing early warnings of hydrometeorological extremes, including floods and droughts, is fundamental in a sustainable society. In terms of observations, extremes pose a particular challenge because they are, by definition, rare events and therefore poorly sampled in many observational records. In terms of AI training data, besides being enough extensive to include an adequate number of extreme events, it is crucial that their full magnitude is accurately represented. Not fulfilling these requirements can lead to statistical bias, overfitting, and inadequate representation of the governing processes, ultimately constraining the predictive skill of trained models^7,8. The recent emergence of physics-informed AI, e.g. Physics-Informed Neural Networks (PINN), is expected to safeguard process representation, reducing the risk of “butterfly effects”, by forcing the models to obey fundamental physical laws^9,10.

In hydrometeorology, the primary sources of large training datasets for AI development include remote sensing products, reanalysis datasets, and outputs from process-based hydrological models (a non-exhaustive overview of major global datasets commonly used for AI training is provided in Box 1). Although these data sources span a wide range of spatial scales, from point measurements to gridded fields or catchment-integrated estimates, and temporal resolutions, from sub-hourly to annual, they share a critical dependency: their accuracy ultimately relies on in situ observations (Fig. 1).

Remote sensing products used in AI applications in hydrometeorology are predominantly derived from satellite-based Earth Observations (EO), which provide spatially continuous estimates of key environmental variables such as precipitation, soil moisture, and flood extent⁷. Although EO is often perceived as raw imagery, it results from complex modelling and retrieval processes that convert sensor measurements into geophysically meaningful variables¹¹. Numerous studies have highlighted the essential role of in situ observations in this transformation, particularly for calibrating, bias-correcting, and validating EO-derived products^12,13, including cutting-edge missions such as the Surface Water and Ocean Topography (SWOT)^14,15. Its potential to observe water elevation with an accuracy of a few dm could not have been determined without access to high-quality in-situ measurements.

Reanalysis products assimilate meteorological observations into numerical weather prediction models, thereby providing physically consistent estimates of a wide range of variables beyond those directly observed. Although in situ measurements are incorporated into the reanalysis process, comparisons with independent in situ observations frequently reveal systematic biases^13,16. Consequently, a range of post-processing approaches has been developed to improve reanalysis accuracy through adjustment towards in situ measurements^17,18.

Hydrological model outputs used in AI applications provide temporally continuous simulations of hydrological variables over extended periods, most notably streamflow, but also soil moisture, evapotranspiration, and related fluxes. In process-based hydrological modelling, model structure, parameters, and their interactions are calibrated against in situ observations to reduce uncertainty and improve predictive skill^19,20. Parameter estimation and model calibration have therefore been central topics in hydrological research for several decades²¹. Consistent with findings across other data sources, large and spatially extensive in situ datasets are generally required to achieve robust regional-scale predictions at high spatial resolution.

Collecting and providing high-quality in situ observations is a complex and highly resource-intensive activity. Initially, careful planning is required to identify locations that are climatologically representative, that have the required infrastructure for access and data transfer, and that are sufficiently secure to avoid theft, vandalism, or other forms of disruption. Sensors, sensor networks and data collection platforms are generally expensive to purchase as well as to install. Ensuring long-term functionality requires a stable operational organization with continual funding for regular maintenance as well as for replacing malfunctioning components. Finally, the collected data needs to undergo careful quality assurance and quality control (QA/QC), preferably by a combination of automated processes and human review, which requires additional financial support and stability.

Several open, global in situ databases provide measurements of key hydrometeorological variables, including precipitation, soil moisture, and streamflow (a non-exhaustive overview of major global in situ databases is provided in Box 2). As a consequence of the conditions and resources required, however, the spatial distribution of open in situ observation stations and networks, as well as the lengths of their associated time series, is highly uneven across the globe (Fig. 2). Data from the Global Runoff Data Centre (GRDC; Fig. 2a) illustrate this disparity, ranging from regions with dense networks and long observational records—such as North America, much of Europe, and eastern Australia—to regions characterized by sparse coverage and short time series, including most of Africa, the Middle East, and Southeast Asia²². A broadly similar pattern is evident in the Global Historical Climatology Network (GHCN; Fig. 2b), although with notable regional differences, and in other global in situ databases²³.

In addition, global open hydrometeorological in situ data have experienced a long-term decline, driven both by the closure of monitoring stations and by reduced data sharing^24,25,26. The World Bank 2018 reported that approximately two-thirds of national hydrological networks are in decline, with no evidence of improvement in subsequent years²⁷. Crochemore et al., in their comprehensive compilation of global in situ runoff records, documented a downward trend in available time series beginning in the early to mid-1980s²², a pattern also apparent in the number of stations recorded in the GRDC database (Fig. 2c). Contributing factors include the decommissioning of old stations, limited maintenance, and increasing restrictions on recent data due to commercial or political considerations. A broadly similar decline is observed in meteorological networks, as reflected in the in-situ observations underlying the Global Precipitation Climatology Centre (GPCC) products²⁸. While the temporal patterns vary among contributing datasets, the overall trend indicates reduced station coverage since the mid-1980s, with the Global Historical Climatology Network (GHCN) showing a pronounced global decline beginning around 2005 (Fig. 2d).

The rapid progress of AI in hydrometeorology has been possible only because decades of in situ measurements provided the foundation for the calibration and validation of hydrological and meteorological models, as well as for the adjustment of remote sensing products. However, the pronounced global inequality and long-term decline in in situ observations pose a threat to the reliability of future AI systems. Data gaps translate directly into predictive gaps, potentially limiting scientific advancement and the equitable distribution of its benefits.

Reduced availability of in situ measurements increases the risk that both remote sensing and reanalysis products become more biased and uncertain, particularly for extreme events. Consequently, AI tools trained on these data may exhibit lower predictive skill, potentially leading to inaccurate assessments of climate change impacts and reduced reliability of forecasting and early warning systems. While physics-informed AI is expected to generate more robust and less big-data demanding models, high-quality in-situ data remain indispensable for e.g. accurate parameterization and validation. For hydrological predictions, data-driven models depend on observed hydrological records as training targets, restricting direct model development to gauged locations²⁹. Predictions at ungauged locations therefore rely on the model’s capacity to generalize across space, a process that benefits from diverse training datasets encompassing a wide range of catchments and hydrological responses^30,31. Indeed, studies have shown that models trained on datasets with multiple response types achieve improved generalization, underscoring the critical need for broader in situ observation coverage at the global scale³².

There is a clear risk that the emerging potential of AI in hydrometeorology will be constrained by the declining availability and quality of both open and professional in situ data. Rather than enabling technological “leapfrogging,” the unequal global distribution of observations—particularly for extremes and climate change impacts—may perpetuate inequities in knowledge and predictive capacity. Even developed countries are likely to be affected, as the progressive reduction in available data will gradually erode the performance and reliability of AI-based tools.

Safeguarding the future of AI in hydrometeorology requires urgent and coordinated action. For sustainable AI progress, it is absolutely essential that the actors advocating, funding, and developing AI solutions for hydrometeorology fully realize and acknowledge the critical role of high-quality in situ measurements as well as the resources required to obtain them. Governments and funding agencies should prioritize the maintenance and modernization of existing in situ observation networks to prevent further degradation of the observational foundation on which AI applications depend. This support must be considered as a natural, integrated, and inevitable part of AI investments.

International collaboration should focus on expanding observational coverage in data-scarce regions, ensuring that low- and middle-income countries (LMICs) can equitably benefit from AI-driven advances in hydrometeorology³³. Strengthening data-sharing practices and open-access policies is also essential, enabling the integration of heterogeneous data sources into robust training datasets. Aligning these efforts with global initiatives such as the UN Early Warnings for All and AI for Good platforms could help leverage existing momentum toward a more resilient and equitable global observing system³⁴.

Innovative and resource-efficient management of observation networks is equally important. This includes optimizing network design under financial constraints by preserving information content while removing redundant sensors with limited added value^26,35. Complementary data sources from non-professional and so-called opportunistic sensors should also be explored and integrated where appropriate^36,37. Examples include 3D printed automatic weather stations³⁸, low-cost river-level sensors³⁹, as well as online private weather stations, which are already used operationally in several European countries⁴⁰.

In parallel, scientific and community-driven initiatives can help counter the ongoing decline of traditional observation networks. In particular, the compilation and open dissemination of large-sample datasets—including in situ observations from diverse sources—are critical for benchmarking, reproducibility, and the development of AI-based methods. A prominent example is the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset⁴¹, which initially focused on daily streamflow in the conterminous United States but has since been extended to multiple countries⁴².

Without decisive action, the potential of AI to improve predictions of floods, droughts, and other hydrometeorological hazards is unlikely to be realized—not because of methodological limitations, but due to the continued erosion of the observational data required to support these approaches.