Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative weather for improved crop model simulations

Yuji Saikai

TL;DR

The paper addresses the gap in weather input preparation for long-term crop yield simulations by introducing a task-specific generative weather model based on a lightweight neural network with dilated causal convolutions. It formalizes a probabilistic weather-generation framework with a factorized likelihood and trains it on historical data to sample future weather conditioned on the past. In two representative crop scenarios, the generated weather consistently improves yield predictions compared with conventional subsequence-based methods, as measured by the mean and standard deviation of absolute errors. The authors provide comprehensive technical details, trained models, and data to enable crop modellers to apply the method with APSIM or other crop models.

Abstract

Accurate and precise crop yield prediction is invaluable for decision making at both farm levels and regional levels. To make yield prediction, crop models are widely used for their capability to simulate hypothetical scenarios. While accuracy and precision of yield prediction critically depend on weather inputs to simulations, surprisingly little attention has been paid to preparing weather inputs. We propose a new method to construct generative models for long-term weather forecasts and ultimately improve crop yield prediction. We demonstrate use of the method in two representative scenarios -- single-year production of wheat, barley and canola and three-year production using rotations of these crops. Results show significant improvement from the conventional method, measured in terms of mean and standard deviation of prediction errors. Our method outperformed the conventional method in every one of 18 metrics for the first scenario and in 29 out of 36 metrics for the second scenario. For individual crop modellers to start applying the method to their problems, technical details are carefully explained, and all the code, trained PyTorch models, APSIM simulation files and result data are made available.

Generative weather for improved crop model simulations

TL;DR

The paper addresses the gap in weather input preparation for long-term crop yield simulations by introducing a task-specific generative weather model based on a lightweight neural network with dilated causal convolutions. It formalizes a probabilistic weather-generation framework with a factorized likelihood and trains it on historical data to sample future weather conditioned on the past. In two representative crop scenarios, the generated weather consistently improves yield predictions compared with conventional subsequence-based methods, as measured by the mean and standard deviation of absolute errors. The authors provide comprehensive technical details, trained models, and data to enable crop modellers to apply the method with APSIM or other crop models.

Abstract

Accurate and precise crop yield prediction is invaluable for decision making at both farm levels and regional levels. To make yield prediction, crop models are widely used for their capability to simulate hypothetical scenarios. While accuracy and precision of yield prediction critically depend on weather inputs to simulations, surprisingly little attention has been paid to preparing weather inputs. We propose a new method to construct generative models for long-term weather forecasts and ultimately improve crop yield prediction. We demonstrate use of the method in two representative scenarios -- single-year production of wheat, barley and canola and three-year production using rotations of these crops. Results show significant improvement from the conventional method, measured in terms of mean and standard deviation of prediction errors. Our method outperformed the conventional method in every one of 18 metrics for the first scenario and in 29 out of 36 metrics for the second scenario. For individual crop modellers to start applying the method to their problems, technical details are carefully explained, and all the code, trained PyTorch models, APSIM simulation files and result data are made available.
Paper Structure (23 sections, 9 equations, 3 figures, 8 tables)

This paper contains 23 sections, 9 equations, 3 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Model
  3. Weather generation problem
  4. Likelihood
  5. Loss function
  6. Dilated causal convolution
  7. Neural network architecture
  8. Construction of the training data
  9. Experiments
  10. Scenarios
  11. Location and original dataset
  12. Crop model specifications
  13. Architecture specifications
  14. Other details
  15. Results
  16. ...and 8 more sections

Figures (3)

  • Figure 1: $\textcolor{myorange}{\blacktriangle}$ represents an output and $\color{lightgray}\bullet$ represents a padded 0. Each output value is computed based on $T-1$ values, which may consist of past observations $\color{myblue}\bullet$, outputs from the preceding time steps $\color{myblue}\blacktriangle$, and padded 0s $\color{lightgray}\bullet$. Only two receptive fields, the one for the first output and the other for the last output are explicitly drawn. Notice that the last output is the only one whose receptive field does not contain any padded 0s.
  • Figure 2: Causal convolution using a filter of size 2 without dilation. The receptive field of $\textcolor{myorange}{\blacktriangle}$ node is four $\textcolor{myorange}{\bullet}$ at the input layer.
  • Figure 3: Dilated causal convolution using a filter of size 2 with dilation factors of 1, 2, 4 and 8. The receptive field of $\textcolor{myorange}{\blacktriangle}$ node is eight $\textcolor{myorange}{\bullet}$ at the input layer. Without increasing the number of hidden layers or trainable parameters, dilated causal convolution achieves a larger receptive field.