The Paradox of AI Progress and the Decline of Hydrometeorological Observation Networks: An Infrastructure Perspective

Introduction

Hydrometeorological disasters such as floods and droughts are among the most destructive natural disasters worldwide. AI is regarded as a key technology for enhancing prediction, monitoring, and early warning capabilities, and has made significant progress in recent years in areas such as extreme weather forecasting and runoff simulation. However, the effectiveness of AI models heavily depends on the quality and representativeness of training data. Currently, the global in-situ hydrometeorological observation network is facing a decline, a trend that stands in sharp contrast to the flourishing of AI, potentially fundamentally limiting the practical utility of AI technology, especially posing systemic infrastructure risks for developing countries.

The Dependence of AI on In-Situ Observations

AI models build predictive capabilities by learning complex relationships from training data. In the hydrometeorological field, the main data sources include remote sensing products, reanalysis datasets, and outputs from process-based hydrological models. Although these data appear independent, their accuracy ultimately relies on in-situ observations: remote sensing products require calibration and bias correction through field measurements; reanalysis products, while assimilating various observational data, still need independent field measurements for validation and further adjustment; and parameter estimation and calibration of hydrological models have long depended on dense in-situ observation networks for river flow, precipitation, and other variables.

Special Requirements for Extreme Events

Extreme events are inherently rare, requiring training data that not only span a long temporal range but also accurately record the full magnitude of extreme values. In-situ observation stations are the only means to provide high-quality records of extreme events. If the network is insufficient, AI models will face statistical bias, overfitting, and inadequate representation of processes, thereby limiting predictive skill. Recent proposals of physics-informed neural networks (PINNs) can introduce physical constraints but cannot fully replace data representativeness.

Decline and Imbalance of Global Observation Networks

Establishing and maintaining high-quality in-situ observation networks is a resource-intensive activity. However, many regions around the world are witnessing funding cuts, equipment aging, and management deficiencies in hydrometeorological observation stations. The problem is particularly acute in low- and middle-income countries (LMICs), which already have sparse observations, and further network decline exacerbates data gaps. Ironically, these are precisely the countries that most hope to achieve a technological “leapfrog” using AI to enhance disaster warning and resource management at lower cost. Yet insufficient data makes the foundation for AI applications in these regions weak, potentially widening the technology gap.

Case Study: Limitations of Satellite Missions

Take the SWOT satellite altimetry mission as an example. Its ability to precisely measure water surface elevation (with decimeter-level accuracy) relies heavily on high-quality field measurements for validation. Without sufficient ground stations, the uncertainties of satellite products cannot be effectively constrained, and the results from AI models trained on them will be unreliable. This highlights the necessity of complementarity between remote sensing and in-situ observations.

Re-examination from an Infrastructure PerspectiveFrom the perspective of infrastructure investment and engineering projects, the hydrometeorological observation network is a typical public infrastructure characterized by long-term, high fixed costs, and high social returns. In the Global South, the inadequacy of such infrastructure directly limits climate resilience building, water resource management, and agricultural planning. Tools such as PPP models and multilateral development bank financing could be used to support network construction, but observation stations often lack direct commercial returns and have long been marginalized.

Impact on Engineering Capital Flows

International engineering firms and investment institutions, when participating in climate adaptation projects (such as flood control levees, irrigation systems, early warning systems), require reliable hydrological data to support design decisions. The decline of observation networks means increased data risks in the early stages of projects, which may raise project financing costs or lead to overly conservative designs. In the long run, the lack of local data will inhibit capital flows to infrastructure projects in climate-vulnerable regions.

Recommendations and Outlook

To resolve this paradox, a multi-pronged approach is needed:

1. International coordination and funding commitments: Institutions such as the WMO and the World Bank should promote a special fund for the modernization of global observation stations, with a particular tilt towards LMICs. 2. Technology integration: Use AI itself to optimize the layout of observation networks (e.g., adaptive sampling, low-cost sensors), but care must be taken to avoid over-reliance on synthetic data. 3. Public-private partnerships: Explore making hydrometeorological data a digital public good to attract private sector participation in operations (e.g., through data service transactions). 4. Open source and standardization: Promote data sharing agreements and format unification to lower barriers to accessing observation data.

AI progress should not be built on "quicksand." Without simultaneously strengthening on-site observation as a foundational infrastructure, the AI revolution in hydrometeorology will eventually face a ceiling. For the Global South, this is even more a window of development opportunity—infrastructure investment and AI capacity building must advance in parallel to truly unleash the long-term dividends of climate resilience.