# A Comprehensive Modeling Approach for Crop Yield Forecasts using AI-based Methods and Crop Simulation Models

**Authors:** Renato Luiz de Freitas Cunha, Bruno Silva, Priscilla Barreira Avegliano  
**arXiv:** [2306.10121](https://arxiv.org/abs/2306.10121)  
**Date:** 2023-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.10121v1/

## Key points

- Comprehensive approach for yield forecasting in crop management decision-making
- Combination of data-driven solutions, crop simulation models, and model surrogates
- Calibration of crop simulation models at scale
- Surrogate model for faster execution
- Neural network-based approach for efficient risk assessment
- Distributions for yield prediction instead of point estimates
- Use of daily temperature, precipitation, solar irradiation, growing degree days (GDD), and soil-related data as features
- Incorporation of CHIRPS dataset for monthly precipitation data and satellite/weather station information for rainfall deviation estimation
- Monthly air temperature data from reanalysis datasets used
- Inclusion of GDD as a feature to account for temperature influence on crops
- Updated model architecture with noise generation layers, LSTM recurrent layers, dense fully-connected layers, and concatenation layers.
- High yield correlation predictions close to 91% in the data-driven modeling approach.
- Crop simulation modeling architecture achieved an error rate of 6%
- Proposed crop simulation model surrogate performed predictions almost 100 times faster with similar accuracy levels.

## Summary

This paper proposes a comprehensive approach for yield forecasting in crop management decision-making. The approach combines data-driven solutions, crop simulation models, and model surrogates to support multiple user-profiles and needs. The researchers have developed a solution to calibrate crop simulation models at scale, a surrogate model of a crop simulation model for faster execution, and a neural network-based approach for efficient risk assessment. The paper extends the existing model by providing distributions for yield prediction instead of point estimates. This allows users to perform risk assessment using probability distribution functions in the decision-making process for farming management. The model incorporates daily temperature, precipitation, solar irradiation, growing degree days (GDD), and soil-related data as features. The researchers used the CHIRPS dataset as a source for monthly precipitation data and incorporated satellite and weather station information to estimate rainfall deviation from climatologies. Monthly air temperature data was provided by reanalysis datasets from era5. GDD was included as a feature to account for the influence of temperature on different crops. The updated model architecture includes noise generation layers, Long Short-Term Memory (LSTM) recurrent layers, dense fully-connected layers, and concatenation layers. The model achieved high yield correlation predictions close to 91% in the data-driven modeling approach. The crop simulation modeling architecture achieved an error rate of 6%, while the proposed crop simulation model surrogate performed predictions almost 100 times faster with similar accuracy levels. Overall, this refined summary highlights the novel contributions of this paper in combining data-driven solutions with crop simulation models and model surrogates to improve yield forecasting in crop management decision-making processes.