Multi-stage growth-aware maize yield prediction using graph neural networks

Maize is the most widely grown cereal on Earth, and predicting how much grain a field will produce before harvest has long been one of agriculture’s most stubborn problems. Farmers, breeders, and policymakers all depend on yield estimates, whether to plan irrigation and fertilizer schedules, to decide which hybrid lines deserve advancement in a breeding program, or to anticipate shortfalls in national food supplies. Yet most prediction models share a fundamental weakness: they typically rely on plant traits measured at a single growth stage, usually close to maturity when the season’s outcome is already largely written. A research team in China has now unveiled a new artificial intelligence framework that breaks with that convention, drawing on phenotypic and environmental data collected across multiple developmental stages to forecast final maize yield earlier, more accurately, and more robustly than conventional approaches.

The framework, called Yield-Graph, is described in a study published in Theoretical and Applied Genetics by Jiahui Wang, Yong Zhang, Yuqing Zhang, Xinglin Piao, and Aiwen Wang of Beijing University of Technology, together with Bo Li of the University of Macau and Xiangyu Zhao and Kaiyi Wang of the Beijing Academy of Agriculture and Forestry Sciences. The work is the product of a collaboration spanning computer science and applied agricultural research, and it reflects a growing conviction among computational biologists that the machinery of modern graph-based machine learning is uniquely suited to the messy, incomplete, and highly interconnected data that agricultural science actually produces.

The conceptual foundation of the study is simple but consequential. A maize plant is not a static object; it is a developmental trajectory. Traits such as plant height, flowering time, leaf architecture, and ear characteristics emerge sequentially, and each stage contributes differently, and sometimes unpredictably, to the final number of kernels that end up in the grain tank at harvest. A model that looks only at late-season traits captures the endpoint of that trajectory but misses the earlier signals that reveal how the season is unfolding. A model that looks only at early traits may predict too coarsely. By modeling multiple traits across different developmental stages, all oriented toward the same target of final yield, Yield-Graph can quantify the stage-specific contributions of each trait and demonstrate that meaningful yield prediction is feasible well before the crop matures.

The technical heart of the framework consists of two graph-based components that address the two most persistent obstacles in agricultural prediction: missing data and complex, higher-order relationships. Field phenotyping is notoriously incomplete. Sensors fail, weather interrupts measurement campaigns, and some traits simply cannot be recorded for every plant in every plot. Traditional machine learning pipelines either discard incomplete samples, wasting hard-won field data, or fill gaps with crude statistical imputations that ignore the structure of the data. Yield-Graph instead constructs a bipartite graph structure that links samples to trait observations, allowing the model to impute missing trait values at each developmental stage by leveraging the relational structure of the dataset itself. Rather than treating each missing value in isolation, the graph embedding approach learns from the patterns of similarity and relatedness among samples, reconstructing absent measurements in a way that is consistent with the surrounding network of observations. This strategy draws on a lineage of graph representation learning methods developed for handling missing data in general machine learning contexts, adapted here to the specific structure of multi-stage crop phenotypes.

The second component tackles a subtler challenge. Samples in agricultural datasets, whether individual plants, plots, or genotype-environment combinations, are not related to one another only in pairs. A group of plants sharing the same environment, the same management regime, and similar phenotypes form a higher-order relationship that cannot be fully expressed as a collection of pairwise connections. Mathematics has a natural tool for such relationships: the hypergraph, a generalization of a graph in which a single edge can connect any number of nodes. Yield-Graph employs a hypergraph attention mechanism to capture these high-order sample relationships. Attention mechanisms allow the model to learn, during training, which groups of related samples and which features within them are most informative for predicting yield, weighting their contributions adaptively rather than treating all relationships as equally important. This allows the framework to model environment-trait interactions of a complexity that ordinary feature-based models cannot easily represent, effectively embedding an awareness of the agricultural context into the architecture of the network itself.

Benchmark experiments reported in the study show that Yield-Graph matches the top-tier predictive accuracy of exhaustively optimized tree-based models, the gradient-boosting ensembles such as XGBoost and LightGBM that have dominated tabular prediction tasks in agriculture in recent years. That parity is significant, because tree ensembles typically demand extensive hyperparameter tuning and careful feature engineering to reach their peak performance, whereas the graph-based framework achieves comparable accuracy while natively accommodating incomplete data and higher-order structure. In other words, the new method reaches the same destination by a route that is better matched to the realities of field data.

Perhaps more important than raw accuracy are the robustness results. The study reports that Yield-Graph exhibits strong robustness across growth stages, meaning that its predictive skill does not collapse when measurements are taken earlier or later in the season, a property essential for genuinely early yield forecasting. It also shows high adaptability to regional variations, retaining accuracy when applied across geographically distinct production environments where soil, climate, and management practices differ. And it generalizes effectively across datasets, a quality that speaks to the framework learning transferable representations of the relationship between growth, environment, and yield rather than memorizing the quirks of a single trial. For a tool intended to be deployed in precision agriculture, where new seasons, new regions, and new crop varieties are the norm rather than the exception, such generalization is the difference between an academic curiosity and a practical instrument.

The broader context of this work is a field in rapid transformation. Crop yield prediction has historically proceeded along several parallel tracks: statistical models linking weather variables to historical yields, mechanistic crop simulations that model plant physiology day by day, satellite-based approaches that estimate yield from vegetation indices, and, more recently, machine learning models that combine genomic, environmental, and phenotypic predictors. Each approach has contributed, but each also faces limitations, particularly when data are sparse, noisy, or incomplete, as they invariably are in real agricultural settings. Graph neural networks have begun to make inroads, with previous work applying spatial-temporal graph learning to regional yield forecasting from geospatial data. What distinguishes Yield-Graph is its focus on plant-level, multi-stage phenotypic data and its explicit design for the incompleteness and higher-order structure that characterize such data, positioning it as a complement to remote sensing and environmental modeling rather than a replacement.

The implications for breeding programs are particularly intriguing. Modern genomic prediction methods already integrate environmental covariates to model genotype-by-environment interactions, and recent work has assembled large “enviromic” datasets to improve the accuracy of yield plasticity prediction in maize. Yield-Graph adds a temporal dimension to this picture: by tracking how traits expressed at seedling, vegetative, and reproductive stages each contribute to the final outcome, breeders gain a tool for early-stage selection, identifying promising lines mid-season rather than waiting for harvest data. This could compress breeding cycles and reduce the cost of field trials, in which thousands of candidate lines must be evaluated each year. The study’s authors also note the framework’s applicability across crop species, suggesting that the underlying methodology is not tied to maize specifically, though maize, with its well-characterized developmental stages and global economic importance, is the natural proving ground.

The team has made the code publicly available through a GitHub repository, and the heading dataset used in the analysis is accessible through Figshare, reflecting a commitment to reproducibility that will ease adoption and further development by other groups. The datasets generated during the study are otherwise available from the corresponding author on reasonable request. The research was supported by the National Key Research and Development Program of China, the Open Project of the National Innovation Center for Digital Seed Industry, and the Beijing Academy of Agricultural Artificial Intelligence and Robotics.

As climate variability intensifies and global demand for grain continues to climb, the value of seeing a season’s outcome before it happens will only grow. Recent studies have documented increasingly nonlinear relationships between daily temperature extremes and agricultural yields, and new generations of climate and crop models suggest that climate impacts on agriculture are emerging sooner than previously expected. In that environment, forecasting tools that are accurate, early, and resilient to imperfect data are not luxuries but necessities. Yield-Graph’s authors frame their contribution as a scalable solution for precision agriculture and intelligent crop management, and the benchmark evidence they present makes a credible case. Whether the framework will scale from research plots to the vast, heterogeneous realities of commercial farming remains to be tested, but the study marks a clear step in the convergence of graph machine learning and agricultural science, a convergence that promises to change how, and how early, humanity predicts the harvest that feeds it.