Core Concepts
A new method using neural networks can generate long-term weather forecasts that significantly improve the accuracy and precision of crop yield predictions compared to the conventional approach of using historical weather records.
Abstract
The paper proposes a new method to construct generative models for long-term weather forecasts and ultimately improve crop yield prediction. The key aspects are:
Scenario 1 (Single-year production):
- The method generates 1000 sets of 365-day weather sequences for 2021, 2022 and 2023, and uses them as inputs to APSIM crop models for wheat, barley and canola.
- Compared to the conventional method of using historical weather records, the generative weather model outperformed the conventional method in every one of 18 metrics for mean and standard deviation of yield prediction errors.
Scenario 2 (Three-year production using crop rotations):
- The method generates 1000 sets of 1095-day weather sequences for 2021-2023, and uses them as inputs to APSIM crop models for 6 different three-year crop rotations.
- The generative weather model outperformed the conventional method in 29 out of 36 metrics for mean and standard deviation of yield prediction errors.
The paper explains the technical details of the generative weather modeling approach, including the likelihood specification, loss function, dilated causal convolution, and neural network architecture design. It also discusses the flexibility and trade-offs of the method compared to the conventional approach.
Stats
The average daily absolute difference in solar radiation between the generated and true weather is around 4.8 MJ/m^2.
The average daily absolute difference in minimum temperature is around 2.9°C.
The average daily absolute difference in maximum temperature is around 3.0°C.
The average daily absolute difference in rainfall is around 2.6 mm.
Quotes
"Accurate and precise crop yield prediction is invaluable for decision making at both farm levels and regional levels."
"To overcome the infeasibility of NWP and provide a practical method for weather input preparation, we propose neural network models that can be trained on historical records at locations of interest and generate desired numbers of future weather values."
"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."