New AI Framework Predicts Global Crop Yields Months Before Harvest

When a pandemic shuts down borders, wildfires sweep across farmland, or war erupts in one of the world’s breadbaskets, the question that haunts governments and food agencies alike is deceptively simple: how much food will the world’s fields actually produce this year? A new study published in Nature Food offers the most comprehensive answer yet to that question. An international team of researchers led by Ziyue Chen of Beijing Normal University has built an integrative framework that fuses official yield statistics with a rich suite of satellite observations and complementary geospatial data, using random forest machine learning models to estimate yields of major crops in every production country on Earth. Crucially, the framework was stress-tested against three of the most disruptive events of the past decade: the COVID-19 pandemic, the catastrophic Australian wildfires, and the war in Ukraine.

The challenge the researchers set out to solve is one that has frustrated agricultural scientists for years. Global crop yield estimation is extraordinarily difficult because the planet’s farmland is staggeringly diverse. Crop phenology varies enormously from one region to the next, with wheat sown in autumn in some countries and spring in others. Environmental conditions differ across climates, from irrigated river deltas to rain-fed plains. Agricultural practices range from precision farming in Europe and North America to smallholder systems elsewhere. Existing monitoring systems tend to be regional, crop-specific, or dependent on ground data that simply does not exist at scale in many countries. The new framework sidesteps these obstacles by learning directly from the relationships between what satellites observe and what farmers ultimately harvest, country by country and crop by crop.

Technically, the framework is built on a foundation of multi-source remote sensing. The team drew on vegetation indices such as the normalized difference vegetation index and the enhanced vegetation index, both well-established proxies for the greenness and vigor of growing crops. They added daytime and nighttime land surface temperature records, temperature, precipitation and evapotranspiration data from climate reanalysis products, and dynamic land cover information from the Dynamic World dataset, which maps land use at ten-meter resolution in near real time. Crop type distributions came from the SPAM v2020 global dataset, planting and harvesting calendars from the Crop Calendar Dataset maintained by the Center for Sustainability and the Global Environment, and administrative boundaries from the FAO’s Global Administrative Unit Layers. Official yield and harvested area statistics for the four focal crops were obtained from FAOSTAT.

These heterogeneous data streams were then fed into random forest models, an ensemble machine learning technique introduced by Leo Breiman in 2001 that builds hundreds of decision trees on random subsets of the data and averages their predictions. Random forests are well suited to this problem because they handle large numbers of correlated predictor variables, capture nonlinear relationships between growing conditions and final yield, and resist overfitting when trained on noisy real-world data. The models were trained to translate the seasonal trajectory of satellite-observed growing conditions into a final yield figure for each country, allowing the same underlying machinery to operate across dramatically different agricultural systems.

The true test of any forecasting system is how it performs when the world goes wrong, and the researchers selected three disruptions that could hardly be more different in character. The first was the COVID-19 pandemic, a truly global shock that disrupted supply chains, labor availability and trade flows in virtually every country simultaneously. The second was the Australian wildfire season of 2019 and 2020, a regional catastrophe in which fires of unprecedented intensity burned millions of hectares, threatening croplands and degrading the very satellite signals that monitoring systems depend on, since smoke and scorched earth complicate the interpretation of vegetation indices. The third was the war in Ukraine, which erupted in 2022 in a country that ranks among the world’s leading exporters of wheat, maize and sunflower, sending shockwaves through global grain markets and raising fears of food shortages far beyond the conflict zone.

The results, reported in four main figures in the paper, are striking. Across the globe, the framework achieved satisfactory accuracy in estimating yields for the major crops, and its performance was strongest precisely where it matters most for food security: in the major crop-producing countries with developed agricultural techniques. When the researchers compared estimated yields against official statistics for the top ten producing countries in 2020, the agreement was robust, demonstrating that a single unified framework could match or approach the accuracy of systems tailored to individual crops or regions. Analyses of how model performance varied with different sets of input features also revealed which satellite-derived signals carried the most information at different points in the growing season, offering a practical guide for building early warning systems.

Perhaps the most consequential finding concerns timing. When croplands were not severely affected by events such as wars or wildfires, the framework enabled crop yield estimates months before harvest. In Australia, the team demonstrated early estimation of yields for four crops sown in both 2019 and 2020, showing that the models could track growing conditions and converge on accurate yield predictions well before combines entered the fields. In Russia and Ukraine in 2022, the framework produced estimates of harvest yields and, importantly, quantified the errors in those early estimates, giving decision-makers a realistic picture of both the expected harvest and the uncertainty surrounding it during one of the most volatile periods in modern grain trade history.

The implications for global food security are difficult to overstate. Expectations of crop yields in major production countries strongly influence export policies, and those policies can cascade through world markets with remarkable speed. During the pandemic, grain export restrictions imposed by some countries risked pushing low- and middle-income importers toward food insecurity, and research has shown that even trade policy announcements alone can increase price volatility in global food commodity markets. A trusted, months-ahead yield estimate could give governments and international agencies the lead time to coordinate responses, adjust trade flows, and prevent panic-driven policies from amplifying a harvest shortfall into a hunger crisis. Satellite-based harvest forecasting has already been shown to trigger cross-hemispheric production responses, and a framework that works in all production countries extends that capability to the entire planet.

The study also arrives at a moment of mounting pressure on the global food system. Climate change is reducing yields of major crops across multiple independent estimates, threatening crop diversity at low latitudes, and intensifying climate shocks in smallholder agriculture across sub-Saharan Africa and the Asia-Pacific region. Invasive pests, soil degradation and wildfire activity add further layers of risk. Against this backdrop, a scalable early estimation framework is not a luxury but a form of infrastructure, comparable in importance to weather forecasting or epidemiological surveillance. The authors emphasize that their research provides a methodological reference for global yield estimation that can support timely crop trade policies and reduce food security risks.

Notably, the team has made the work transparent and reproducible. The underlying datasets span openly available resources, from FAOSTAT statistics and NASA’s land products to Copernicus climate data and the Armed Conflict Location and Event Data project, and the source code is publicly available on GitHub. That openness matters because the framework’s value will ultimately depend on how quickly agencies, researchers and policymakers can adapt and deploy it. As disruptions of every kind, from pandemics to conflicts to climate extremes, continue to test the resilience of the world’s food supply, the ability to know months in advance how the harvest is shaping up, anywhere on Earth, may prove to be one of the most quietly powerful tools of the coming decade.

Random forest models occupy a distinctive niche among machine learning approaches for agricultural prediction. Unlike deep neural networks, which typically demand vast training datasets, random forests can perform well with comparatively modest samples of country-level observations, making them attractive for a problem where labeled yield data exist only as annual statistics. Their ensemble structure also yields measures of variable importance, which likely underpinned the study’s analysis of how different input features contributed to early estimation skill as the growing season progressed.

The choice of vegetation indices reflects decades of remote-sensing research. NDVI, computed from red and near-infrared reflectance, exploits the fact that healthy chlorophyll-rich canopies absorb red light strongly while scattering near-infrared radiation. EVI improves on this in dense canopies, where NDVI tends to saturate. Pairing these optical signals with nighttime land surface temperature is particularly informative, since minimum temperatures during sensitive growth stages can sharply constrain final grain numbers even when vegetation appears green.

The three test cases also probe different failure modes of monitoring systems. The pandemic tested robustness to widespread socioeconomic disruption without direct biophysical damage to crops. The Australian fires tested resilience to atmospheric aerosols and burned landscapes that corrupt optical observations. The Ukraine war tested performance amid active conflict, where ground-truthing is impossible and official statistics themselves become uncertain, which is precisely why the Armed Conflict Location and Event Data were incorporated as an input layer.

Accuracy was strongest in countries with advanced agricultural systems, a reminder that satellite-based estimation inherits the quality of the statistics it learns from. Extending reliable early estimates to smallholder-dominated regions, where yield variability is high and data sparse, remains the central challenge for the next generation of global crop monitoring.

Subject of Research: Early estimation of global crop yields using satellite remote sensing and machine learning under large-scale disruptions

Article Title: An integrative framework for early estimation of global crop yields demonstrated under large-scale disruptions

Article References: Chen, Z., Wang, Y., Yan, X., Kwan, M.-P., Yang, L., Wang, Q., Zhu, Q., Yu, Q., Feng, Z., Gao, B., Zhang, C., Fu, Y., Hu, J., Li, M., & Wang, Q. (2026). An integrative framework for early estimation of global crop yields demonstrated under large-scale disruptions. Nature Food. https://doi.org/10.1038/s43016-026-01418-w

Image Credits: AI Generated

DOI: 10.1038/s43016-026-01418-w

Keywords: crop yields, remote sensing, random forest, machine learning, food security, COVID-19, Australian wildfires, Ukraine war, satellite data, global agriculture, early yield estimation, Nature Food