PaperlayerView exact revision on arXiv

Source-linked AI summary

A CNN-RNN Framework for Crop Yield Prediction

Saeed Khaki, Lizhi Wang, Sotirios V. Archontoulis

arXiv:1911.09045v2cs.LGq-bio.QMstat.ML

TL;DR

Crop-yield prediction must account for interacting genotype, environmental, and management factors, while genotype data may be unavailable. The paper combines CNNs and RNNs to model environmental and management data, achieving lower validation error than the tested alternatives for corn and soybean yields across the Corn Belt. The framework also analyzes variable importance over time and supports prediction in untested environments.

  • Problem

    Crop-yield prediction is difficult because genotype, environmental, and management factors interact, while genotype data may be unavailable.

  • Method

    The paper combines CNNs, fully connected layers, and LSTM-based RNNs to process environmental and management data and model temporal yield dependencies.

  • Results

    The CNN-RNN outperformed LASSO, random forest, and DFNN, with validation RMSE approximately 9% of average corn yield and 8% of average soybean yield.

  • Takeaways & Limitations

    The framework predicted yields in untested environments and estimated variable importance by time period for weather, soil, and management inputs.

  • Takeaways & Limitations

    The study lacked publicly available genotype data to complement its yield, management, weather, and soil datasets.

Abstract

from arXiv · show

Crop yield prediction is extremely challenging due to its dependence on multiple factors such as crop genotype, environmental factors, management practices, and their interactions. This paper presents a deep learning framework using convolutional neural networks (CNN) and recurrent neural networks (RNN) for crop yield prediction based on environmental data and management practices. The proposed CNN-RNN model, along with other popular methods such as random forest (RF), deep fully-connected neural networks (DFNN), and LASSO, was used to forecast corn and soybean yield across the entire Corn Belt (including 13 states) in the United States for years 2016, 2017, and 2018 using historical data. The new model achieved a root-mean-square-error (RMSE) 9% and 8% of their respective average yields, substantially outperforming all other methods that were tested. The CNN-RNN have three salient features that make it a potentially useful method for other crop yield prediction studies. (1) The CNN-RNN model was designed to capture the time dependencies of environmental factors and the genetic improvement of seeds over time without having their genotype information. (2) The model demonstrated the capability to generalize the yield prediction to untested environments without significant drop in the prediction accuracy. (3) Coupled with the backpropagation method, the model could reveal the extent to which weather conditions, accuracy of weather predictions, soil conditions, and management practices were able to explain the variation in the crop yields.

INTRODUCTION

Crop yield prediction is difficult because genotype, environmental factors, management practices, and their interactions jointly affect yields. The paper motivates a deep-learning framework that learns representations from environmental and management data while combining CNN and RNN capabilities.

  • Crop yield varies with genotype, environment, and management practices, whose interactions make accurate prediction difficult.
  • Accurate yield prediction supports food-production planning, farmer decisions, and evaluation of hybrids in untested locations.
  • Environmental factors such as weather can have complex nonlinear effects that are difficult to estimate accurately.
  • Deep learning learns multiple levels of nonlinear representations directly from data without handcrafted features.
  • The proposed hybrid model combines CNNs for array-formatted data with RNNs that capture dependencies in sequential data.

DATA

The study combines yield, management, weather, and soil datasets for crop-yield modeling. Weather and soil inputs are processed into learned features before being combined and passed to the recurrent component.

  • The analyzed data comprised yield performance, management, weather, and soil datasets, without publicly available genotype data.
  • Yield observations covered corn and soybean across 1,176 and 1,115 counties, respectively, within 13 Corn Belt states from 1980 to 2018.
  • Missing soil and management values were imputed using means from comparable counties or years.
  • Daily weather data were averaged weekly, reducing dimensionality by a 365:52 ratio before neural-network modeling.
  • The proposed model combines CNNs, fully connected layers, and RNNs to process crop-yield inputs.
  • W-CNN captures temporal weather dependencies, while S-CNN captures spatial dependencies in soil measured at different depths.

RNN

The RNN component models crop-yield dynamics across years using LSTM cells and historical yield information. It indirectly represents genetic improvement despite the absence of genotype data.

  • The RNN captures crop-yield time dependencies motivated by long-term yield increases and unavailable genotype data.
  • The model uses LSTM cells to capture temporal dynamics associated with genetic improvement in crop yield.
  • For year t, k LSTM cells predict county yield using information from years t −k through t.
  • RNN inputs include average yield, management data, and features extracted from weather and soil by CNN and fully connected layers.
  • During testing, prior-year average yield and predicted weather can substitute for unavailable target-year inputs.

DESIGN OF EXPERIMENTS

The experiments train the CNN-RNN model and compare it with RF, DFNN, and LASSO using historical Corn Belt yield data. Figure 1 supplies long-term yield trends, while Figure 2 depicts the model’s unrolled temporal structure.

  • The CNN-RNN used a 5-year recurrent time length, with four convolutional layers in both weather and soil CNNs.
  • The model was trained with Xavier initialization, mini-batch SGD, Adam optimization, ReLU activations, and a linear output layer.
  • The CNN-RNN model was compared with random forest, DFNN, and LASSO as alternative prediction methods.
  • Figure 1 plots average corn and soybean yields across the Corn Belt from 1980 to 2018 in bushels per acre.
  • Figure 2 unrolls the CNN-RNN structure and identifies weather, soil, yield, predicted yield, and management inputs at each time step.

RESULTS

Across 2016–2018 validation years, the CNN-RNN model outperformed RF, DFNN, and LASSO for corn and soybean yield prediction, with low errors across most counties. Its performance was attributed to modeling temporal weather and yield dependencies, soil spatial dependencies, and nonlinear environmental effects.

  • The CNN-RNN model outperformed RF, DFNN, and LASSO across all measures for all three validation years.
  • 9% and 8% were the validation RMSEs as proportions of the respective average yields for corn and soybean.The RMSE unit was bushels per acre.
  • The model’s performance was attributed to capturing yearly yield dependencies, temporal weather dependencies, soil-depth spatial dependencies, and nonlinear environmental effects.
  • The CNN-RNN model maintained consistent performance for both crops across all validation years.
  • Prediction errors were consistently low for most counties in the 2018 corn and soybean error maps.Black counties indicated unavailable ground-truth yields; map errors were measured in bushels per acre.
  • The CNN-RNN model approximately preserved some distributional properties of ground-truth corn and soybean yields for 2018.

Feature Selection

The study uses guided backpropagation to estimate how weather, soil, and management variables affect yield across time and identifies compact feature subsets that preserve prediction accuracy.

  • Feature importance estimation: Guided backpropagation ranks input variables by positive gradients that maximize targeted-neuron activation.The analysis was performed separately for corn and soybean yield prediction.
  • Feature importance estimation: Estimated effects were normalized within weather, soil, and management groups to make variables comparable.The analysis included weather components, soil conditions at the surface and different depths, and management practices.
  • Weather effects: Solar radiation was the most sensitive weather factor for corn, while snow was the least sensitive.Radiation showed sensitive periods around weeks 15 and 30, whereas snow was more important before and after the growing season.
  • Weather effects: Precipitation became increasingly important during the soybean grain-fill period from weeks 30 to 40.
  • Soil effects: Soil variables sensitive to both corn and soybean predictions were linked to soil water and nitrogen supply.The paper notes that explaining every sensitive soil factor is beyond its scope.
  • Feature reduction: Selecting the 50% or 75% most important features did not significantly reduce CNN-RNN prediction accuracy.Feature selection used 2017 data and evaluated predictions on 2018 data.

Importance Comparison Between Environment and Management Practices

The component comparison evaluates weather, soil, and management information separately, finding weather and soil more informative for yield prediction than management practices.

  • Comparison design: The comparison used CNN-RNN models based separately on weather, soil, or management data, alongside an average-yield benchmark.
  • Management effects: Planting dates had greater effects on soybean yield prediction than on corn yield prediction.
  • Results: Weather-only CNN-RNN performance was comparable to soil-only CNN-RNN performance for both corn and soybean.
  • Results: Weather and soil explained more yield variation than management practices in the tested comparisons.

Generalization Power of the CNN-RNN Model

The study tests whether CNN-RNN yield predictions transfer to locations excluded from training, using five-fold cross-validation on 2018 yields.

  • Evaluation design: Locations were randomly excluded from 1980–2017 training data and then used as untested locations for 2018 prediction.The procedure used k-fold cross-validation to estimate generalization power.
  • Evaluation design: Five-fold cross-validation was used to evaluate CNN-RNN generalization for corn and soybean yield prediction.Table 6 reports the performance using RMSE in bushels per acre.
  • Results: Prediction accuracy did not deteriorate considerably relative to the corresponding standard results.The authors interpret this as successful generalization to untested locations.

Yield Prediction Using Predicted Weather Data

Because weather is unknown before the prediction period, the study evaluates CNN-RNN yield prediction using predicted rather than perfect weather data.

  • Motivation: Weather prediction is an unavoidable component of crop yield prediction because weather is unknown a priori.
  • Evaluation design: For Iowa’s 2018 predictions, 2017 June–September weather was used as predicted weather for the corresponding 2018 interval.The study compared RMSE and predicted state-average yield under perfect and predicted weather data.

CONCLUSION

The CNN-RNN framework accurately predicted corn and soybean yields across the Corn Belt while extending prediction toward explaining variable importance and improving model interpretability.

  • The model accurately predicted corn and soybean yields across the entire Corn Belt using environmental data and management practices.
  • The methodology moved beyond prediction by providing variable importance by time period, while addressing the black-box property of deep learning models.
  • The hybrid CNN-RNN significantly outperformed LASSO, random forest, and DFNN.CNNs capture temporal weather and spatial soil dependencies, while RNNs capture yield trends associated with breeding and management improvement.
  • The model successfully predicted yields in untested environments, supporting its use in future yield prediction tasks.
  • Feature selection with backpropagation estimated the individual effects of weather, soil, and management variables and when they became important.

SUPPLEMENTARY MATERIALS

The supplementary materials are available online and the cited resources include datasets, source code, and foundational methods relevant to crop-yield prediction and deep learning.

  • The article’s supplementary material is available online through the Frontiers in Plant Science article page.
  • The cited resources include the gSSURGO soil database and Daymet daily surface weather data for North America.
  • The references include prior work on recurrent neural networks, convolutional networks, random forests, LASSO, and crop-yield prediction.
  • Source code for the CNN-RNN yield-prediction study is available through the cited GitHub repository.
Loading 1911.09045v2…