Simple Last-Year Baseline Beats Fancy Graph Models at Predicting Vegetation Productivity

Graph neural networks have become one of the most fashionable tools in environmental science, promising to capture the invisible web of connections that links rivers, climates, and ecosystems across a landscape. But a new study from researchers at Qilu Normal University in Jinan, China, delivers a sobering reality check. In a paper published in Earth Science Informatics, Xinyu Wang, Xianzhi Wang, Duanguang Cao, and Yanying Wang show that when it comes to forecasting vegetation productivity across hundreds of counties in the Yellow River Basin, a deceptively simple baseline — last year’s value — outperformed every sophisticated graph-based model they tested. The finding challenges a growing assumption in geospatial machine learning: that encoding spatial connectivity into a model automatically buys predictive power.

The team’s target was net primary productivity, or NPP, the amount of carbon that plants fix from the atmosphere each year. NPP is a fundamental gauge of ecosystem health, and satellite products such as NASA’s MOD17A3HGF.061 dataset, derived from the MODIS instrument on the Terra satellite, now provide annual estimates at a resolution of 500 meters across the globe. For their analysis, the researchers aggregated these satellite observations to the county level, reconstructing annual NPP for 382 counties spanning the Yellow River Basin between 2003 and 2022. This region is a natural laboratory for such work: it encompasses the Loess Plateau, where decades of large-scale revegetation have pushed ecosystems close to sustainable water-resource limits, and its vegetation dynamics respond to both climatic swings and intense human management.

The methodological heart of the paper is what the authors call a lag-controlled residual benchmark. The idea is elegant in its simplicity. Because vegetation productivity exhibits strong temporal persistence — this year’s NPP in a county is highly correlated with last year’s — any model that predicts NPP directly can look impressive simply by echoing the previous year. To strip away that advantage, the researchers first built a baseline model using only the lag-1 value, meaning the NPP recorded one year earlier. They then trained their graph models not to predict NPP itself, but to predict only the residual: the portion of this year’s productivity that last year’s value fails to explain. If spatial connectivity genuinely carries additional information, the graph models should be able to squeeze signal out of those residuals. If they cannot, the apparent skill of complex spatial models is an illusion created by temporal persistence.

Four graph architectures entered the contest, each encoding a different hypothesis about how counties are connected. The Euclidean-Static graph linked neighboring counties by straight-line distance, the most common and least informative assumption. The Hydro-Static graph connected counties that share hydrological relationships through the river network, reflecting the intuition that water flows, and the ecosystems along it, tie upstream and downstream regions together. The Hydro-Dynamic graph went further, allowing the strength of those hydrological connections to change over time, for instance as rainfall and drought conditions shift the coupling between counties. Finally, an identity-skip graph served as a control, essentially testing whether any learned connectivity mattered at all. Each model was trained to predict the residual NPP beyond the lag-1 baseline, and the experiments were repeated across 20 random seeds to guard against the luck of random initialization, with an independent test period from 2020 to 2022 held out entirely from training.

The results were unambiguous, and for advocates of graph learning, uncomfortable. The lag-1 baseline alone achieved a root mean square error of 0.0404, with an R-squared of 0.8977, meaning it explained nearly ninety percent of the variance in annual county-level NPP using nothing but the previous year’s value. The four graph models, despite their elaborate connectivity structures, produced root mean square errors between 0.0497 and 0.0506 — consistently and substantially worse than the baseline they were supposed to improve upon. In other words, when the models were forced to explain only what temporal persistence could not, they failed to find meaningful signal in the spatial structure of the basin.

The researchers also probed whether the hydrological graphs at least held an edge under particular conditions. They stratified their evaluation by wetness regime, comparing dry, normal, and wet years, and by subbasin, reasoning that dynamic hydro-climatic connectivity might matter most where water stress is severe or where riverine coupling is strongest. No group showed a robust dynamic advantage. A paired Wilcoxon test, a nonparametric statistical comparison suited to repeated runs across seeds, found that the Hydro-Dynamic graph did not significantly outperform the simpler symmetric Hydro-Static graph, with a p-value of 0.0696. The extra complexity of time-varying connectivity bought nothing statistically detectable.

To rule out the possibility that the problem lay with graph neural networks specifically rather than spatial modeling in general, the team also benchmarked conventional residual-target spatial models. Geographically weighted regression, a classical technique that fits locally varying relationships across space, performed best among these alternatives, achieving a root mean square error of 0.0443. That is better than the graph models, and it shows that the classical spatial econometric toolkit retains real value, but it still fell short of the lag-1 baseline. The pattern held across every family of methods: at the annual county scale, increasingly complex representations of spatial connectivity did not translate into stable predictive gains beyond the temporal benchmark.

Why does this matter beyond the Yellow River Basin? The implications reach into a broader debate about how environmental machine learning is evaluated. Surveys of spatio-temporal graph neural networks document explosive growth in their application to traffic, climate, hydrology, and ecology, yet rigorous comparisons against simple temporal baselines remain rare. When a target variable is strongly autocorrelated in time — as annual productivity, temperature, or streamflow often are — a naive persistence forecast can be remarkably hard to beat, and models that blend spatial and temporal information can appear skillful while contributing little beyond what persistence already provides. The lag-controlled residual framework offers a reproducible diagnostic for separating these two sources of skill, and it costs almost nothing to implement: train a lag baseline, compute residuals, and test whether spatial structure explains any of what remains.

The study also carries a practical message for anyone modeling vegetation in a changing climate. Annual NPP in this basin is dominated by memory — carried by standing biomass, soil conditions, root systems, and slow-recovering vegetation cover — rather than by year-to-year information flowing between neighboring counties through river networks or climatic teleconnections. If researchers want to detect genuine spatial coupling, they may need finer temporal resolution, where seasonal anomalies and extreme events such as droughts propagate through landscapes, rather than annual aggregates in which persistence swamps everything else. Drought-legacy research has shown that ecosystem responses to water stress can echo for years, and capturing such effects may require the very kind of dynamic modeling the graph architectures were meant to provide — but only once the temporal baseline is properly controlled.

For a field racing to publish ever-larger and more elaborate models, the paper is a reminder that the most important competitor is often the simplest one. The authors, who report no external funding and no competing interests, frame their contribution not as an attack on graph learning but as a benchmarking discipline: a way to ask, honestly and reproducibly, whether spatial connectivity adds anything at all. In this case, the answer was no — and that negative result, rigorously established across seeds, architectures, wetness regimes, and subbasins, may prove more influential than any positive one. As graph-based geospatial modeling matures, studies like this one will help distinguish genuine advances in our understanding of environmental systems from elaborate repackaging of temporal inertia.

Subject of Research: Benchmarking graph-based geospatial models against a lag-1 temporal baseline for predicting county-level vegetation net primary productivity in the Yellow River Basin

Article Title: A lag-controlled residual benchmark for graph-based geospatial modelling of vegetation productivity

Article References: Wang, X., Wang, X., Cao, D., & Wang, Y. (2026). A lag-controlled residual benchmark for graph-based geospatial modelling of vegetation productivity. Earth Science Informatics, 19(10), Article 169. https://doi.org/10.1007/s12145-026-02232-5

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02232-5

Keywords: graph neural networks, net primary productivity, Yellow River Basin, temporal persistence, residual benchmark, geospatial modelling, MODIS remote sensing, hydrological connectivity, geographically weighted regression, ecosystem forecasting, spatial machine learning, model evaluation