🤖 machine learning

Reconstructing GRACE Terrestrial Water Storage with Spatio-Temporal Graph Neural Networks: An Application to South America

This paper presents a novel deep learning framework using a multi-variate time series graph neural network (MTGNN) to reconstruct monthly GRACE-like terrestrial water storage anomalies for South America back to 1940, achieving high accuracy and basin-scale performance comparable to state-of-the-art methods while requiring significantly fewer predictors by effectively modeling both local hydrological coupling and large-scale teleconnections.

Original authors: Lukas Arzoumanidis, Lara Johannsen, Klara Middendorf, Annette Eicker, Youness Dehbi

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lukas Arzoumanidis, Lara Johannsen, Klara Middendorf, Annette Eicker, Youness Dehbi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Rewinding the Water Clock

Imagine the Earth's water cycle (rain, rivers, soil moisture, and groundwater) as a giant, invisible bank account. Scientists have been keeping a very accurate ledger of this account since 2002 using special satellites called GRACE. These satellites don't take photos; they act like ultra-sensitive scales, weighing the Earth to see where water is moving.

However, there's a problem: The ledger only goes back to 2002. For scientists trying to understand long-term climate change, droughts, or floods, two decades isn't enough. They need to see what happened in the 1940s, 50s, and 60s.

This paper presents a "time machine" built with Artificial Intelligence. The authors created a system that uses old weather data to reconstruct (guess with high accuracy) what that water bank account looked like from 1940 to today, filling in the missing decades before the satellites existed.

The Problem with Old Methods

Before this paper, scientists tried to fill in the missing years using standard math or simple AI.

The Old Way: Imagine trying to predict the weather in every single city in South America by looking at that city alone. You'd look at the rain in Rio and guess the water storage in Rio, then look at the rain in Buenos Aires and guess the storage there.
The Flaw: Water doesn't live in a vacuum. Rain in the Amazon affects rivers that flow hundreds of miles away. A drought in one place is often linked to a flood in another due to giant wind patterns. The old methods treated these cities as isolated islands, ignoring the connections between them.

The Solution: The "Traffic Map" AI

The authors decided to borrow a tool from a completely different field: City Traffic Prediction.

The Analogy: Imagine a traffic AI designed to predict jams. It knows that if a car slows down in one neighborhood, the traffic jam will ripple down the road to the next neighborhood a few minutes later. It treats the road network as a graph (a map of connected dots).
The Transfer: The authors realized that water behaves exactly like traffic.
- Nodes (Dots): Instead of traffic sensors, they used 1,120 grid squares covering South America.
- Edges (Lines): Instead of roads, they mapped how water and weather connect these squares.
- The Flow: Just as a traffic jam moves down a highway, a rainstorm or drought moves through the atmosphere and river systems, affecting distant regions with a delay.

They took this "Traffic AI" (called a Spatio-Temporal Graph Neural Network) and taught it to understand water instead of cars.

How They Built the Map

The most clever part of their work was how they drew the "roads" (connections) between the grid squares.

Distance: They connected squares that are physically close (like neighbors).
Weather Twins: They also connected squares that are far apart but have similar weather patterns (like two cities that always get rain at the same time, even if they are on opposite sides of the continent).

This created a Hybrid Map that understands both local geography and giant global weather patterns (like El Niño).

The Results: How Good Was the Guess?

They tested their "Time Machine" against the real satellite data from 2002 to 2023 to see if it could guess the past correctly.

The Score: It got a 94% match when looking at the average water levels across entire river basins. That is incredibly accurate.
The Details: It successfully recreated the "fingerprints" of major climate events, like the massive droughts and floods caused by the El Niño and La Niña events in 2015 and 2020.
The Efficiency: Other AI models needed to look at 16 or 20 different weather variables (temperature, humidity, solar radiation, etc.) to get a good score. This new model got almost the same score using only 3 variables: Rain, Evaporation, and Runoff. It's like solving a complex puzzle using only three pieces instead of twenty.

Where It Struggles

The paper is honest about its weaknesses. The model works best in wet, rainy areas (like the Amazon) where the water cycle is strong and predictable.

The Weak Spot: In dry, arid regions (like parts of Argentina and Bolivia) where humans pump a lot of groundwater, the model is less accurate. This is a problem for all current models, not just this one, because human activity is hard to predict with just weather data.

The Takeaway

The authors didn't just build a better calculator; they proved that AI tools built for city traffic can be repurposed to understand the Earth's climate. By treating the planet as a connected network rather than a collection of isolated spots, they successfully rewound the clock on water storage data, giving scientists a much longer history to study how our climate is changing.

They have also made their code public, allowing others to try this "traffic map" approach on other parts of the world.

Technical Summary: Reconstructing GRACE Terrestrial Water Storage with Spatio-Temporal Graph Neural Networks

Problem Definition
Terrestrial Water Storage (TWS), which integrates snow, soil moisture, surface water, and groundwater, is a critical indicator of climate variability and human impact on the water cycle. While the Gravity Recovery and Climate Experiment (GRACE) and its successor GRACE-FO provide the only direct, globally consistent observations of TWS changes, their record is limited to the period starting in 2002. This duration is insufficient for multi-decadal climate analyses, such as identifying trends, detecting tipping points, or attributing variability to drivers like the El Niño–Southern Oscillation (ENSO).

Existing reconstruction methods attempt to extend the GRACE record backward using statistical models, physically based hydrological models, or machine learning (e.g., CNNs, LSTMs). However, most of these approaches treat grid cells as spatially independent or rely on implicit spatial context, failing to explicitly model the inherently networked nature of the water cycle, where hydrological connectivity and atmospheric teleconnections couple distant regions.

Methodology
This paper proposes a deep learning application using a Multivariate Time Series Graph Neural Network (MTGNN) to reconstruct monthly GRACE-like TWS anomalies (TWSA) back to 1940. The study focuses on South America as a test region within a globally designed workflow.

Architecture Transfer: The authors adapt the MTGNN architecture, originally developed by Wu et al. for urban traffic forecasting, to the domain of satellite geodesy. The structural analogy is drawn between traffic congestion propagation on road networks and water/atmospheric moisture propagation along river systems and circulation patterns.
Input and Target: The model learns the relationship between daily ERA5 meteorological forcing variables (precipitation $P$ , evapotranspiration $E$ , and runoff $R$ ) and monthly GRACE TWSA observations. The input is a 30-day window of daily flux data, and the target is the monthly TWSA value for each grid cell.
Graph Construction: A key methodological adaptation is the replacement of the MTGNN's learned adjacency matrix with a static, interpretable hybrid adjacency matrix. This matrix encodes spatial dependencies through two components:
1. Geodesic Proximity: A distance-decay Gaussian kernel based on physical distance between grid cells.
2. Lagged Climate Correlations: The maximum-lag Pearson correlation between time series of climatic variables at different nodes, capturing teleconnections where one region's signal leads or trails another's.
  These are combined via a convex combination ( $\alpha=0.6$ for climate similarity) to form the graph structure.
Training Regime: The model is trained on the GRACE/GRACE-FO period (April 2002–December 2023) using a sequential train/validation/test split. Due to the limited number of training samples (183 months), the design employs strong regularization (dropout, weight decay) and moderate model capacity.

Key Contributions

Methodological Transfer: The paper demonstrates that a graph architecture from urban mobility forecasting can be successfully transferred to gravimetric satellite data, provided the graph structure is adapted to encode domain-specific knowledge (teleconnections and hydrological coupling).
Static Hybrid Graph: The authors document the design decision to replace learned graph structures with a static, domain-informed hybrid adjacency matrix. This choice enhances interpretability and stability, avoiding the convergence issues associated with jointly optimizing latent graphs on heterogeneous meteorological inputs.
Efficiency and Parsimony: The proposed model achieves statistical competitiveness using only three input variables ( $P, E, R$ ), whereas established baselines (GTWS-MLrec, RM-REC, GRAiCE) utilize significantly larger predictor sets (ranging from 6 to over 25 variables).
Comprehensive Evaluation: The study provides a multi-stage evaluation including quantitative metrics, seasonal and spatial error analysis, and specific case studies of the 2015/16 El Niño and 2020/21 La Niña events.

Results
Evaluated against GRACE/GRACE-FO data (2002–2023) over South America:

Performance Metrics: The reconstruction achieves a grid-cell Pearson correlation of 0.69 and a basin-mean correlation of 0.94, with a near-zero bias (−0.004 m).
ENSO Events: The model successfully reproduces the spatial fingerprints of major ENSO events, capturing the contrasting drought/wet patterns of the Amazon and La Plata basins during El Niño and La Niña, though it slightly underestimates the amplitude of these extremes.
Comparison with Baselines: When compared to GTWS-MLrec, RM-REC, and GRAiCE, the MTGNN model is statistically competitive at the basin scale (correlation within 0.025 of the best baseline). While it shows slightly lower performance at the single-grid-cell level in arid regions compared to some baselines, it achieves this with a fraction of the input data requirements.
Limitations: Like other reconstruction methods, the model exhibits reduced accuracy in arid and semi-arid regions (e.g., parts of Bolivia, Paraguay, Argentina, and Chile) and areas with strong human influence, where hydroclimatic processes are complex.

Significance and Claims
The paper claims that this work represents the first application of spatio-temporal graph deep learning to TWS reconstruction that explicitly encodes hydroclimatic teleconnections in the graph topology. The significance lies in demonstrating that the analogy between traffic flow and water transport is not merely conceptual but can be directly exploited in graph-based architectures to model global geophysical fields.

The authors emphasize that the choice between learned and prescribed graphs is a critical modeling decision. By using a static, interpretable graph, the model offers a topology-aware perspective on historical water storage that complements existing per-cell approaches. The study concludes that the graph-based approach offers a favorable accuracy-per-variable trade-off, lowering data acquisition costs and making the learned mapping easier to attribute to physical drivers. The authors note that the current performance gap in arid regions is a community-wide challenge and outline future extensions involving input parameter expansion (e.g., adding temperature) and physics-informed training via the terrestrial water balance equation.