Prediction of municipality-level winter wheat yield based on meteorological data using machine learning in Hokkaido, Japan.
This study analyzed meteorological constraints on winter wheat yield in the northern Japanese island, Hokkaido, and developed a machine learning model to predict municipality-level yields from meteorological data. Compared to most wheat producing areas, this island is characterized by wet climate ow...
Guardado en:
| Autores principales: | , , , , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
Public Library of Science (PLoS)
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/153cb3ff3c33401ab65957d9f4b67f8e |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | This study analyzed meteorological constraints on winter wheat yield in the northern Japanese island, Hokkaido, and developed a machine learning model to predict municipality-level yields from meteorological data. Compared to most wheat producing areas, this island is characterized by wet climate owing to greater annual precipitation and abundant snowmelt water supply in spring. Based on yield statistics collected from 119 municipalities for 14 years (N = 1,516) and high-resolution surface meteorological data, correlation analyses showed that precipitation, daily minimum air temperature, and irradiance during the grain-filling period had significant effects on the yield throughout the island while the effect of snow depth in early winter and spring was dependent on sites. Using 10-d mean meteorological data within a certain period between seeding and harvest as predictor variables and one-year-leave-out cross-validation procedure, performance of machine learning models based on neural network (NN), random forest (RF), support vector machine regression (SVR), partial least squares regression (PLS), and cubist regression (CB) were compared to a multiple linear regression model (MLR) and a null model that returns an average yield of the municipality. The root mean square errors of PLS, SVR, and RF were 872, 982, and 1,024 kg ha-1 and were smaller than those of MLR (1,068 kg ha-1) and null model (1,035 kg ha-1). These models outperformed the controls in other metrics including Pearson's correlation coefficient and Nash-Sutcliffe efficiency. Variable importance analysis on PLS indicated that minimum air temperature and precipitation during the grain-filling period had major roles in the prediction and excluding predictors in this period (i.e. yield forecast with a longer lead-time) decreased forecast performance of the models. These results were consistent with our understanding of meteorological impacts on wheat yield, suggesting usefulness of explainable machine learning in meteorological crop yield prediction under wet climate. |
|---|