Summarize Publication
Summarize Publication:
What has been previously done, what are the major conclusion
Major Conclusions
Environmental components are of significant importance in predicting crop yield. Previous models took limited variables into account, leading to a need for a ML model with more layers to gain a more comprehensive understanding of the which variables affect crop yield. Deep learning based predictive modeling works better than the previously done process based models.
What has been previously done?
Previously, models utilized limited variables, limiting the scope of the prediction model.
Ran two LSTM models and these models overcome the error-backflow problems of an RNN. The two models are stacked and LSTM and Temporal Attention Model. These models output a yearly seed yield. They ran these models based on maturity group and Genotype Cluster.
How they did it (made the model):
Long Short-Term Memory is overarching model, gradient-based neural network https://direct.mit.edu/neco/article/9/8/1735/6109/Long-Short-Term-Memory
Adapted LSTM to include Temporal Attention to account for changing monthly, weekly, and daily weather data over time
Different variable inputs: just weather variables, as well as weather variables, maturity group, and genotype cluster
Greedy Search Method - Utilized to empirically determine the most influential weather variable on seed yield prediction considering data of both the northern and southern U.S. regions.
Utilized Three Evaluation Metrics: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination or R-squared (R2) score. These were computed after inverting the applied scaling to have forecasts and the actual values in the original scale.
For comparison of the empirical results, we used two baseline models: Support Vector Regression with Radial Basis Function kernel (SVR-RBF) and least absolute shrinkage and selection operator (LASSO) regression.
Findings and Results:
“The finding of minimum surface temperature as the most significant weather variable (when all variables included, including cluster ID-genotype) suggests that nighttime temperatures play a larger role in yield prediction than previously suggested [64].”
However, the model remains somewhat limited in its ability to generate genotype-specific yield predictions due to the limited complexity of relationships which can be modeled using LSTM, and a lack of genomic information on each genotype.