TL;DR
This paper demonstrates that incorporating fertilizer data significantly improves statistical models' ability to predict crop yields for maize, rice, and wheat. By including fertilizer as a predictor alongside climate variables, these models can better capture yield variability, offering more accurate projections than climate-only models.
What Problem Does This Paper Address?
Accurately predicting crop yields is crucial for global food security, agricultural planning, and understanding the impacts of climate change. Many statistical models used for this purpose have historically relied solely on climate variables (temperature, precipitation) as predictors. However, this approach overlooks other significant factors influencing crop growth, such as the application of fertilizers. The authors aim to address this gap by investigating the extent to which including fertilizer data can enhance the predictive power of statistical crop yield models, particularly for major crops like maize, rice, and wheat. They question whether climate alone is sufficient to explain inter-annual yield variability or if fertilizer use plays a critical, often underestimated, role.
Methods
The researchers employed robust linear regression models, which are less sensitive to outliers than standard least squares methods. They developed separate models for each crop (maize, rice, and wheat) at a grid-cell level. A key aspect of their methodology was the inclusion of annual mean fertilizer data (specifically nitrogen fertilizer) as an additional predictor alongside seasonal rainfall and mean seasonal temperature. Before model fitting, all input data (yields, meteorological data, and fertilizer input) were detrended to focus on explaining inter-annual yield variations rather than long-term trends. This detrending process, using a two-degree polynomial, was chosen to remove gradual changes attributed to factors like breeding, technology, or environmental drivers, thereby isolating the impact of year-to-year climate and fertilizer variations. Model performance was evaluated using metrics like Root Mean Square Error (RMSE), standard deviation (SD), and R² values, and statistical significance was assessed.
Datasets
The study utilized two primary climate forcing datasets:
- Princeton dataset: An updated and extended version of the Princeton University Hydroclimatology Group Bias Corrected 59-yr (1948-2006) Meteorological Forcing Dataset.
- WFDEI dataset: The WATCH-Forcing-Data-ERA-Interim dataset (2014).
Yield data was sourced from the UN FAO (FAOSTAT, 2014) and gridded to match the meteorological data resolution. Planting and harvest dates were obtained from the Ag-GRID GGCMI harmonisation project (Elliott et al., 2015). Fertilizer data, specifically annual mean nitrogen fertilizer in kg/ha/yr, was extracted from supplementary data by Lassaletta et al. (2014).
Key Insights and Ideas
- Fertilizer is a critical, often overlooked, predictor: The study strongly suggests that fertilizer use is not merely a secondary factor but a significant driver of crop yield variability, comparable in importance to climate.
- Climate-only models are insufficient: Relying solely on climate variables leads to an incomplete picture of yield dynamics, potentially misattributing variability and underestimating the impact of agricultural practices.
- Statistical models can be powerful with the right inputs: By augmenting simple statistical models with relevant agricultural data like fertilizer use, their predictive accuracy can be substantially improved, making them more robust for both historical reconstruction and future projections.
- Regional variations matter: The impact of fertilizer inclusion varies by region and crop, highlighting the need for spatially explicit modeling that accounts for local agricultural practices.
Key Results
- Improved Yield Forecasts: The inclusion of nitrogen fertilizer as a predictor improved yield forecasts by an average of 18% using the Princeton climate dataset and 23% using the WFDEI climate dataset.
- Reduced Error: Models incorporating fertilizer showed a reduction in RMSE, indicating they better replicated observed yields. For example, the improvements in yield estimates for the Princeton dataset showed an average improvement of 434 kg/ha for maize, 695 kg/ha for rice, and 472 kg/ha for wheat.
- Increased Explained Variability: The R² values increased when fertilizer was included, meaning the models explained a larger fraction of the observed yield variability.
- Statistical Significance: In most regions, models with fertilizer input were found to be statistically significant (p < 0.05), demonstrating the robustness of the findings.
- Regional Improvements: Significant improvements were observed in countries like India, Pakistan, Argentina, Brazil, South Africa, and parts of Europe, particularly for maize and wheat.
Limitations and Future Work
- Nitrogen-only fertilizer data: The study focused solely on nitrogen fertilizer, lacking specific data for phosphorus or potassium. This limitation means that yield increases attributed to nitrogen might, in some cases, be due to other nutrient applications. Future work could involve more detailed models with specific nutrient channels to address this.
- Country-level data resolution: The fertilizer and yield data were primarily at a country level, which was then gridded. This coarse resolution might mask finer-scale variations in fertilizer application and yield.
- Short-term focus of linear models: While effective for inter-annual variability, simple linear models may not be ideal for long-term projections that require accounting for gradual changes in technology, breeding, and environmental drivers (like CO2). The authors suggest combining these methods with forecasts of climate and fertilizer additions for near-term predictions, and using process-based models for longer-term assessments.
- Detrending limitations: The polynomial detrending, while useful for isolating inter-annual variability, does not remove "step changes" in yield data, which could represent abrupt shocks or data artefacts.
- Model applicability to smaller nations: Outputs for smaller nations were noted as less robust due to potential data limitations.
- Future work: The authors suggest combining their statistical modeling approach with climate and fertilizer forecasts for near-term predictions and developing detailed nitrogen schemes within more complex process-based models for longer-term projections.
Summary updated