Comment on essd-2022-423

Yuchuan Luo, Zhao Zhang, Juan Cao, Liangliang Zhang, Jing Zhang, Jichong Han, Huimin Zhuang, Fei Cheng, Jialu Xu, Fulu TaoOpenalex Unknown 2023 · DOI · OpenAlex

Abstract

Accurate and spatially explicit information on global crop yield is paramount for guiding policy-making and ensuring food security. However, most public datasets are at coarse resolution in both space and time. Here, we used data-driven models to develop a 4-km dataset of global wheat yield (GlobalWheatYield4km) from 1982 to 2020. First, we proposed a phenology-based approach to map spatial distributions of spring and winter wheat. Then we determined the optimal grid-scale yield estimation model by comparing the performance of two data-driven models (i.e., Random Forest (RF) and Long Short-Term Memory (LSTM)), with publicly available data (i.e., satellite and climatic data from the Google Earth Engine (GEE) platform, soil properties, and subnational-level census data covering ~11000 political units). The results showed that GlobalWheatYield4km captured 82 % of yield variations with RMSE of 619.8 kg/ha across all subnational regions and years. In addition, our dataset had a higher accuracy (R² ~0.71) as compared with Spatial Production Allocation Model (SPAM) (R² ~ 0.49) across all subnational regions and three years. The GlobalWheatYield4km dataset might play important roles in modelling crop system and assessing climate impact over larger areas.

Introduction

Global food security and climate change mitigation strategies depend heavily on precise agricultural data. However, researchers often face a "resolution vs. duration" trade-off. Existing global datasets are either spatially coarse—such as the Global Dataset of Historical Yields (GDHY) at ~50km resolution—or cover very short temporal windows, like the Spatial Production Allocation Model (SPAM), which provides only intermittent snapshots.

Furthermore, distinguishing between spring and winter wheat is critical due to their distinct growing cycles and climate sensitivities. To address these gaps, the GlobalWheatYield4km dataset was developed, providing the first 40-year (1982–2020) global wheat yield record at a high 4-km resolution.

Methodology

The development of GlobalWheatYield4km utilized a two-stage data-driven framework:

1. Phenology-Based Mapping

To accurately estimate yield, the spatial distribution of wheat must first be identified. The researchers implemented a phenology-based algorithm to distinguish the unique "green-up" and senescence signatures of spring and winter wheat. By analyzing Leaf Area Index (LAI) time series, they created annual distribution maps that account for the specific growth cycles of each variety.

2. Deep Learning Yield Estimation

The study compared two machine learning architectures: Random Forest (RF) and Long Short-Term Memory (LSTM). While RF is robust for tabular data, LSTM—a recurrent neural network—is uniquely suited for agriculture because it can process sequences. This allows the model to "remember" how environmental stressors (e.g., a heatwave during the heading stage versus the ripening stage) cumulatively impact the final harvest.

3. Validation Framework

A "leave-one-year-out" cross-validation approach was used to ensure model generalizability. The models were trained on all available years except one, which was then used for testing across approximately 11,000 subnational political units.

Data Integration

The study integrated a massive volume of multi-source data:

  • Satellite Imagery: NOAA AVHRR NDVI for long-term vegetation health, GLASS LAI for phenology, and GFSAD1KCM for cropland masking.
  • Climate & Soil: TerraClimate (monthly temperature, precipitation, and vapor pressure) and the Harmonized World Soil Database (soil pH, organic carbon, and texture).
  • Ground Truth: Subnational census data from 54 countries, covering roughly 92% of global wheat area and 93% of production.

Key Results and Performance

The GlobalWheatYield4km dataset represents a significant advancement in agricultural monitoring:

  • High Accuracy: The model captured 82% of global yield variations with a Root Mean Square Error (RMSE) of 619.8 kg/ha.
  • Superior Benchmarking: In head-to-head comparisons for 2000, 2005, and 2010, the new dataset achieved an R² of ~0.71, significantly outperforming SPAM’s R² of ~0.49.
  • Low Uncertainty: Nearly 89% of the pixels in the dataset showed an uncertainty (nRMSE) of less than 30%.
  • Geospatial Insights: The data identified Europe as a high-yield leader (e.g., Belgium at ~8,457 kg/ha), while regions like Kazakhstan and Bolivia showed the lowest averages (< 1,000 kg/ha).

Discussion and Limitations

The study demonstrated strong spatial transferability, showing that models trained on data-rich regions (like Russia) could successfully estimate yields in neighboring regions with similar climates but sparse census data (like Kazakhstan).

However, challenges remain. Reliable subnational census data for Africa is scarce, creating gaps in the global map. Additionally, at a 4-km resolution, "mixed pixel" issues persist in areas where wheat is sparsely grown, potentially leading to estimation errors. Future work will focus on utilizing finer-resolution sensors like Sentinel-2 and improving the accessibility of ground-truth census data.

Conclusion

GlobalWheatYield4km provides a high-resolution, long-term benchmark for agricultural research. By leveraging deep learning and phenological mapping, it offers a more accurate and temporally consistent view of global wheat production than previously available, serving as a vital tool for climate impact assessment and food security policy.

Data Availability: The full 1982–2020 dataset is publicly available via Figshare at https://doi.org/10.6084/m9.figshare.10025006.

Summary updated

Abstract

<strong class="journal-contentHeaderColor">Abstract.</strong> Accurate and spatially explicit information on global crop yield is paramount for guiding policy-making and ensuring food security. However, most public datasets are at coarse resolution in both space and time. Here, we used data-driven models to develop a 4-km dataset of global wheat yield (GlobalWheatYield4km) from 1982 to 2020. First, we proposed a phenology-based approach to map spatial distributions of spring and winter wheat. Then we determined the optimal grid-scale yield estimation model by comparing the performance of two data-driven models (i.e., Random Forest (RF) and Long Short-Term Memory (LSTM)), with publicly available data (i.e., satellite and climatic data from the Google Earth Engine (GEE) platform, soil properties, and subnational-level census data covering ~11000 political units). The results showed that GlobalWheatYield4km captured 82 % of yield variations with RMSE of 619.8 kg/ha across all subnational regions and years. In addition, our dataset had a higher accuracy (R<sup>2</sup> ~0.71) as compared with Spatial Production Allocation Model (SPAM) (R<sup>2</sup> ~ 0.49) across all subnational regions and three years. The GlobalWheatYield4km dataset might play important roles in modelling crop system and assessing climate impact over larger areas (DOI of the referenced dataset: <a href="https://doi.org/10.6084/m9.figshare.10025006" target="_blank" rel="noopener">https://doi.org/10.6084/m9.figshare.10025006</a>; Luo et al., 2022b).
BibTeX
@inproceedings{luo2023comment,
  author = {Yuchuan Luo and Zhao Zhang and Juan Cao and Liangliang Zhang and Jing Zhang and Jichong Han and Huimin Zhuang and Fei Cheng and Jialu Xu and Fulu Tao and Yuchuan Luo and Liangliang Zhang and Jing Zhang and Huimin Zhuang and Fei Cheng},
  title = {{Comment on essd-2022-423}},
  booktitle = {Openalex Unknown},
  year = {2023},
  doi = {10.5194/essd-2022-423-rc1},
}
Figure