NeuralCrop: Combining physics and machine learning for improved crop yield predictions

TL;DR

NeuralCrop is introduced, a hybrid GGCM that combines the strengths of an advanced process-based GGCM, resolving important processes explicitly, with data-driven machine learning components, and offers overall improved crop modeling and more reliable yield projections under climate change and intensifying extreme weather conditions.

Abstract

Global gridded crop models (GGCMs) simulate daily crop growth by explicitly representing key biophysical processes and project end-of-season yield time series. They are a primary tool to quantify the impacts of climate change on agricultural productivity and assess associated risks for food security. Despite decades of development, state-of-the-art GGCMs still have substantial uncertainties in simulating complex biophysical processes due to limited process understanding. Recently, machine learning approaches trained on observational data have shown great potential in crop yield predictions. However, these models have not demonstrated improved performance over classical GGCMs and are not suitable for simulating crop yields under changing climate conditions due to problems in generalizing outside their training distributions. Here we introduce NeuralCrop, a hybrid GGCM that combines the strengths of an advanced process-based GGCM, resolving important processes explicitly, with data-driven machine learning components. The model is first trained to emulate a competitive GGCM before it is fine-tuned on observational data. We show that NeuralCrop outperforms state-of-the-art GGCMs across site-level and large-scale cropping regions. Across moisture conditions, NeuralCrop reproduces the interannual yield anomalies in European wheat regions and the US Corn Belt more accurately during the period from 2000 to 2019 with particularly strong improvements under drought extremes. When generalizing to conditions unseen during training, NeuralCrop continues to make robust projections, while pure machine learning models exhibit substantial performance degradation. Our results show that our hybrid crop modelling approach offers overall improved crop modeling and more reliable yield projections under climate change and intensifying extreme weather conditions.

Figures (45)

• Figure 1: Schematic of the NeuralCrop two-stage training framework. It shows how forcings $x_t^{\,n}$, including weather, soil properties, land use, and field management (e.g., crop calendar, tillage, residue, and fertilizer), and previous time-step model states ${y}_{t-1}^{\,m}$ are fed into NeuralCrop at each time step $t$, where $n$ denotes the number of forcing variables and $m$ denotes the number of model states. These model states ${y}_{t-1}^{\,m}$ are fed into the corresponding neural networks of hybrid components, which advance the evolution of model states over time. The new model states ${y}_{t}^{\,m}$ are then fed back into the model to compute next time step. a, In the pre-training stage, NeuralCrop is trained to emulate the behavior of the process-based GGCM LPJmL by minimizing the discrepancies $L(\hat{y}_t^n, y_t^n)$ between NeuralCrop output $\hat{y}_t^n$ and LPJmL output $y_t^n$, given identical weather, soil property, land use, and field management inputs. b, In the fine-tuning stage, the pre-trained NeuralCrop is further trained using observational data from eddy-covariance observation networks, which serve as ground truth.
• Figure 2: Comparison of simulated European wheat yields from NeuralCrop and LPJmL with EU statistics in European wheat regions. a, Time series correlation coefficient between LPJmL simulated wheat yield and EU statistics at the subnational level for the period 2000-2019 (range from $-1$ to $1$, darker green areas indicate stronger positive correlations, and darker purple areas indicate stronger negative correlations). b, Same as (a) but for NeuralCrop. c, The difference of correlation coefficient (i.e., panel b -- panel a), where blue areas indicate regions where NeuralCrop outperforms LPJmL in simulating interannual yield variability, and red areas indicate regions where LPJmL performs better. NeuralCrop performs better (worse) than LPJmL in 61.8% (38.2%) of all subnational regions. d, Boxplots of time series correlation coefficients between simulated wheat yield and EU statistics at the subnational level, aggregated by country for the period 2000–2019, with countries ordered alphabetically by their country code (A–Z). Blue boxes denote LPJmL, and orange boxes denote NeuralCrop. The box boundaries represent the interquartile range (IQR), defined by the first quartile, the median, and the third quartile. The upper and lower whiskers represent the maximum and minimum that are within 1.5 times the interquartile range of the box. The black triangles are the mean values.
• Figure 3: Comparison of simulated corn yields from NeuralCrop and LPJmL with USDA statistics in the US Corn Belt, including nine states (i.e., South Dakota, Minnesota, Iowa, Missouri, Wisconsin, Illinois, Michigan, Indiana, and Ohio). a, Time series correlation coefficient between LPJmL simulated corn yield and USDA statistics at the country level for the period 2000-2019 (range from $-1$ to $1$, darker green areas indicate stronger positive correlations, and darker purple areas indicate stronger negative correlations). b, Same as (a) but for NeuralCrop. c, The difference of correlation coefficient (i.e., panel b -- panel a), where blue areas indicate regions where NeuralCrop outperforms LPJmL in simulating interannual yield variability, and red areas indicate regions where LPJmL performs better. NeuralCrop performs better (worse) than LPJmL in 53.1% (46.9%) of all subnational regions. d, Boxplots of time series correlation coefficients between simulated corn yield and USDA statistics at the country level, aggregated by states for the period 2000–2019. Blue boxes represent LPJmL simulations vs. USDA statistics, and orange boxes represent NeuralCrop simulations vs. USDA statistics. The box boundaries represent the interquartile range (IQR), defined by the first quartile, the median, and the third quartile. The upper and lower whiskers represent the maximum and minimum that are within 1.5 times the interquartile range of the box. The black triangles are the mean values.
• Figure 4: Comparison of simulated yields from NeuralCrop and LPJmL with statistics for European wheat and US Corn Belt across moisture conditions at the subnational level during 2000–2019. a, The boxplots represent the wheat yield anomalies in EU over the period 2000–2019, grouped by the April-June Standardized Precipitation Evapotranspiration Index (SPEI) (Extremely Wet: 1.5 $\le$ SPEI; Very Wet: 1.0 $\le$ SPEI $<$ 1.5; Moderately Wet: 0.5 $\le$ SPEI $<$ 1.0; Normal: -0.5 $\le$ SPEI $<$ 0.5; Moderately Dry: -1.0 $\le$ SPEI $<$ -0.5; Very Dry: -1.5 $\le$ SPEI $<$ -1.0; Extremely Dry: SPEI $<$ -1.5)). Green boxes denote EU statistics, orange boxes denote NeuralCrop, and blue boxes denote LPJmL. The box boundaries represent the interquartile range (IQR), defined by the first quartile, the median, and the third quartile. The upper and lower whiskers represent the maximum and minimum that are within 1.5 times the interquartile range of the box. The black triangles are the mean values . The bars in the lower panel represent the root mean square error (RMSE) for NeuralCrop and LPJmL within each SPEI class. Orange bars denote NeuralCrop, and blue bars denote LPJmL. b, Same as panel a, but for corn yields evaluated against USDA statistics.
• Figure 5: Yield anomalies and model performance for NeuralCrop and LPJmL under the extreme drought years. a, April–June dryness category at the subnational level in the EU in 2018, classified by the April-June Standardized Precipitation Evapotranspiration Index (SPEI) (Normal: -0.5 $\le$ SPEI $<$ 0.5; Moderately Dry: -1.0 $\le$ SPEI $<$ -0.5; Very Dry: -1.5 $\le$ SPEI $<$ -1.0; Extremely Dry: SPEI $<$ -1.5)). b, Wheat yield anomalies in 2018 at the subnational level from EU statistics. c, May–July dryness category at the county level in US Corn Belt in 2012, classified by the May–July SPEI (Normal: -0.5 $\le$ SPEI $<$ 0.5; Moderately Dry: -1.0 $\le$ SPEI $<$ -0.5; Very Dry: -1.5 $\le$ SPEI $<$ -1.0; Extremely Dry: SPEI $<$ -1.5)). d, Corn yield anomalies in 2012 at the county level from USDA statistics. e, Boxplots of simulated wheat yield anomaly bias (i.e., the difference between simulated wheat yield anomalies and statistical yield anomalies) grouped by the dryness classes. Orange boxes denote NeuralCrop and blue boxes denote LPJmL. The box boundaries represent the interquartile range (IQR), defined by the first quartile, the median, and the third quartile. The upper and lower whiskers represent the maximum and minimum that are within 1.5 times the interquartile range of the box. The black triangles are the mean values. f, Same as panel e, but for simulated corn yield anomaly bias (%).