📄 agriculture

Assessing Different Remotely Sensed Wheat Biophysical Variables for Data Assimilation in a Crop Model for Yield Mapping Prediction

This study demonstrates that assimilating Sentinel-2 derived above-ground biomass into the DSSAT CERES-Wheat model significantly improves wheat yield prediction and mapping accuracy at the pixel scale, outperforming schemes based solely on leaf area index or canopy nitrogen content.

Original authors: Riccardo Fazioli, Federica Carucci, Luca Marrone, David Kovács, Davide Cammarano, Stefano Lo Presti, Zhenhai Li, Raffaele Casa

Published 2026-09-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Riccardo Fazioli, Federica Carucci, Luca Marrone, David Kovács, Davide Cammarano, Stefano Lo Presti, Zhenhai Li, Raffaele Casa

Original paper licensed under CC BY 4.0 (https://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

Farmers have long relied on the weather, the soil, and their own experience to guess how much wheat a field will produce. Today, they have a powerful new tool: satellites that can see the crops from space. But simply looking at a satellite image is not enough to predict a harvest. The challenge lies in connecting those distant views of green leaves to the complex, hidden processes happening underground and within the plant itself. To do this, scientists use computer models that simulate how a crop grows, calculating how much water it drinks, how much nitrogen it eats, and how it turns sunlight into grain. These models are like detailed blueprints for a living machine, but they often drift off course because the real world is messy and full of surprises. To fix this, researchers use a technique called data assimilation. Think of it as a navigator constantly checking a map against the actual road signs to correct the route. By feeding real-time observations from satellites into the computer model, scientists can force the simulation to stay aligned with reality, turning a rough guess into a precise forecast.

In a recent study conducted in the wheat fields of central Italy, researchers set out to find the best "road sign" to guide their computer model. They tested three different pieces of information that satellites can measure: the amount of leaf surface area, the total weight of the plant above the ground, and the amount of nitrogen stored in the leaves. The team wanted to know which of these variables, when fed into their simulation, would produce the most accurate prediction of the final harvest. They did not just look at one field; they ran thousands of computer simulations with different weather patterns and soil types to see how the system behaved in theory, and then they tested their ideas on real farms in the Rieti plain during the 2024 growing season.

The results revealed a clear winner. When the researchers fed the model data about the total weight of the plant, known as aboveground biomass, the predictions became remarkably accurate. In their computer simulations, using this single piece of information reduced the error in yield prediction by more than half compared to running the model without any satellite data. When they applied this to the real fields, the improvement was just as striking. The model's estimate of the harvest became much closer to what the farmers actually harvested, with the error dropping from over 1,400 kilograms per hectare to less than 600 kilograms. The spatial patterns of the yield also matched reality much better, capturing the subtle variations across the field rather than just showing a flat, uniform guess.

Surprisingly, the other two variables did not work as well on their own. Feeding the model only the leaf area data, or only the nitrogen content, actually made the predictions worse in some cases, pushing the error higher than if no satellite data had been used at all. The researchers found that while leaf area and nitrogen are important for the plant's health, they are too indirect to reliably predict the final weight of the grain when the satellite measurements contain even small errors. The total weight of the plant, however, is a direct measure of growth that is tightly linked to the final harvest. It acts as a sturdy anchor, holding the model steady even when other inputs are uncertain.

The study also showed that combining all three variables did not necessarily make the result better than using the weight alone. In the real-world tests, adding the less reliable leaf and nitrogen data to the strong weight data did not improve the outcome significantly. This suggests that in the messy reality of a farm, having more data is not always better if that data is noisy. The quality of the measurement matters more than the quantity. By focusing on the most robust variable, the researchers were able to create a high-resolution map of the field's yield, showing exactly where the crop performed well and where it struggled, all based on a limited number of ground samples and satellite images.

This work offers a practical path forward for precision agriculture. It suggests that farmers and agronomists do not need to chase every possible measurement from space to improve their forecasts. Instead, they should focus on the variables that provide the most stable and direct link to the harvest. By using satellite data to track the total weight of the crop, computer models can be tuned to reflect the true state of the field, allowing for better decisions about fertilization and irrigation. The study confirms that while the sky provides a wealth of information, the key to understanding the harvest lies in knowing which piece of that information is the most reliable guide.

Technical Summary: Assessing Different Remotely Sensed Wheat Biophysical Variables for Data Assimilation in a Crop Model for Yield Mapping Prediction

Problem Statement
Process-based crop models, such as DSSAT-CERES-Wheat, are essential tools for agronomic management and yield prediction. However, their accuracy is often limited by uncertainties in input data and their inability to capture the spatial variability inherent in agricultural fields. While data assimilation (DA) has been proposed to integrate remote sensing observations into these models to improve consistency between simulated and observed states, the implications of selecting specific biophysical variables for assimilation remain unclear. Previous studies have predominantly focused on Leaf Area Index (LAI). This research addresses the gap in understanding whether integrating other variables, specifically aboveground biomass and canopy nitrogen content, offers superior performance for yield mapping and whether multivariable schemes outperform single-variable approaches under real-world conditions.

Methodology
The study employed a three-phase framework to evaluate the DSSAT-CERES-Wheat model:

Synthetic Experiment: An in-silico experiment was conducted using 99 factorial simulations combining three soil types, three growing seasons (varying rainfall), and eleven nitrogen fertilization rates. Synthetic "true" values for yield and grain nitrogen were generated, and synthetic observations of LAI, aboveground biomass, and canopy nitrogen content were created by adding 30% random noise to the true state variables. Various assimilation schemes (single and multivariable combinations) were tested against an open-loop baseline using an optimization-based approach (Particle Swarm Optimization).
Field Validation (Real Data): The framework was applied to winter wheat fields in the Rieti Plain, Central Italy, during the 2024 growing season. Sentinel-2 imagery was processed to derive time series of LAI, aboveground biomass, and canopy nitrogen content at 10 × 10 m resolution. These were validated against ground measurements from 10 Experimental Sampling Units (ESUs) across three fields with heterogeneous management practices.
Pixel-Scale Yield Mapping: Using the best-performing multivariable configuration, the assimilation framework was applied at the pixel scale to a 14-hectare field (LIDA15) to generate a high-resolution yield map. This was compared against an interpolated yield map derived from 137 ground sampling points.

The study utilized a trial-and-error pre-calibration to align model parameters with observed phenology and growth before the assimilation step. Performance was assessed using Root Mean Square Error (RMSE), the coefficient of determination ( $R^2$ ), Willmott's Index of Agreement (D-index), and spatial codispersion coefficients.

Key Contributions

Comparative Assessment of Variables: The study systematically compares the efficacy of assimilating LAI, aboveground biomass, and canopy nitrogen content individually and in combination for yield prediction.
Multivariable vs. Single-Variable Analysis: It evaluates whether combining structural (LAI, biomass) and nutritional (nitrogen) variables provides a robust advantage over single-variable assimilation, particularly when transitioning from idealized synthetic data to noisy real-world satellite data.
High-Resolution Yield Mapping: The research demonstrates the practical application of a data assimilation framework to generate spatially explicit, high-resolution (10 × 10 m) yield maps, bridging the gap between point-based ground truth and field-scale model outputs.

Results

Synthetic Conditions: Under idealized conditions, assimilation significantly reduced errors compared to the open-loop model. While LAI alone performed well for yield prediction, multivariable schemes including biomass (CWAD) minimized yield errors further. For grain nitrogen content, the combination of LAI and canopy nitrogen content yielded the best results.
Real-World Field Data: The performance varied significantly based on the quality of the satellite-derived variables.
- Aboveground Biomass: This variable demonstrated the strongest agreement with field measurements ( $R^2 = 0.943$ ) and proved to be the most robust for constraining the model. Assimilating biomass alone reduced yield RMSE by 32% and increased the D-index from 0.30 to 0.71.
- LAI and Canopy Nitrogen: Assimilating LAI or canopy nitrogen content in isolation degraded yield prediction performance compared to the open-loop model, likely due to higher retrieval uncertainties and the indirect relationship of these variables to final yield in the model structure.
- Multivariable Schemes: Combining variables improved results only when biomass was included. The addition of LAI or nitrogen to biomass did not yield systematic improvements over biomass-only assimilation in the real-data context, suggesting that the lower quality/noise of the other variables limited their utility.
Yield Mapping: At the pixel scale, the assimilated yield map showed a realistic spatial pattern and significantly better agreement with observed data than the open-loop simulation. The RMSE decreased from 1415.40 to 576.53 kg DM ha⁻¹, and the spatial codispersion coefficient improved from -0.11 to 0.57, indicating a strong positive spatial correspondence.

Significance and Claims
The paper claims that aboveground biomass is the most effective remotely sensed variable for constraining process-based crop models for yield prediction, outperforming LAI and canopy nitrogen content when derived from Sentinel-2 data in this specific context. The study highlights that the success of multivariable data assimilation is heavily dependent on the quality of the satellite products; while multivariable schemes are theoretically superior, their practical benefit is limited if the additional variables (LAI, nitrogen) introduce significant noise or uncertainty.

The authors conclude that assimilating multiple biophysical variables can improve wheat yield prediction and generate accurate yield maps at the field scale, providing practical guidance for integrating crop models with satellite data in precision agriculture. However, they modestly note that these findings are based on a single growing season and location, and future work is needed to test the framework across different pedoclimatic conditions and to explicitly incorporate observational uncertainty into the assimilation process.

More like this