Fig 1.
Spatial distribution of municipality-level winter wheat yields used in this study.
The number on each municipality (separated by thin lines) indicates the number of years with yield records. Four area groups (separated by thick lines) consist of three or four subprefectures (separated by dashed lines). The inset shows a histogram of the yield (N = 1,516). Yield data are published by Hokkaido Regional Agricultural Administration Office (Ministry of Agriculture, Forestry and Fisheries; URL: https://www.maff.go.jp/hokkaido/toukei/kikaku/sokuho/r2kouhyou.html; accessed on 2021 March 31). The boundaries are published by National Land Information Division, National Spatial Planning and Regional Policy Bureau (https://nlftp.mlit.go.jp/ksj/index.html, accessed on 2021-04-06).
Table 1.
Meteorological statistics of municipalities between 2007 and 2020 in four regional groups.
Fig 2.
Historical trend of municipality-level winter wheat yield in Hokkaido.
Vertical lines represent years with low yields due to a national scale heatwave (2010) and severe rainfall events in summer (2009 and 2018).
Fig 3.
Relationships between the municipality-level yield of winter wheat and total precipitation [Rain, mm], mean value of daily minimum air temperature [Tmin, °C], and total solar irradiance [Irad, MJ m–2] during the grain-filling period (mid June to early July), and mean value of snowpack depth [Snow, cm] in November and March.
Pearson’s correlation coefficients and statistical significance (***: P < 0.001, **: P < 0.01, *: P < 0.05, n.s.: not significant) are shown. Filled symbols in the top-left panel correspond to 2009 and 2018 with severe rainfall events during summer, whereas those in the top-right panel correspond to 2010 with a national-scale heatwave during summer.
Fig 4.
Relationship between actual municipality-level winter wheat yields and predicted ones using machine learning models (N = 1,502).
Years with high yields (2015 and 2019) and low yields (2009, 2010, and 2018) are highlighted. Evaluation metrics are summarized in Table 2.
Fig 5.
Yearly RMSE values of models for winter wheat yield prediction calculated via leave-one-year-out cross-validation.
Table 2.
Predictive performance of selected models and their input periods of meteorological data.
Fig 6.
Relationship between predicted and actual relative yields categorized with a bin size of 20%.
Circles in orange and gray indicate correct and incorrect classification, respectively. Numbers on and sizes of the circle indicate the number of records in each category.
Fig 7.
Accuracy of predicted municipality-level winter wheat yields as a function of an acceptable error.
Each predicted value was regarded as a correct prediction if | (predicted yield) / (actual yield) − 1| was smaller than the acceptable error.
Fig 8.
Performance of machine learning models for winter wheat yield forecast.
Models use meteorological data from early April to different terms as shown on the x-axis. Mean values and standard deviations calculated via one-year-out cross-validation are shown (N = 14).
Fig 9.
Variable importance of a partial least squares model for winter wheat yield prediction.
Mean values and standard deviations calculated via one-year-out cross-validation are shown (N = 14). Loess smoothing curves show trends in the variable importance.