Abstract
Statistical meteorological impact models are intended to represent the impact of weather on socioeconomic activities, using a statistical approach. The calibration of such models is difficult because relationships are complex and historical records are limited. Often, such models succeed in reproducing past data but perform poorly on unseen new data (a problem known as overfitting). This difficulty emphasizes the need for regularization techniques and reliable assessment of the model quality. This study illustrates, in a general way, how to extract pertinent information from weather data and exploit it in impact models that are designed to help decision-making. For a given socioeconomic activity, this type of impact model can be used to 1) study its sensitivity to weather anomalies (e.g., corn sensitivity to water stress), 2) perform seasonal forecasting (yield forecasting) for it, and 3) quantify the longer-term (several decades) impact of weather on it. The size of the training database can be increased by pooling data from various locations, but this requires statistical models that are able to use the localization information—for example, mixed-effect (ME) models. Linear, neural-network, and ME models are compared, using a real-world application: corn-yield forecasting over the United States. Many challenges faced in this paper may be encountered in many weather-impact analyses: these results show that much care is required when using space–time data because they are often highly spatially correlated. In addition, the forecast quality is strongly influenced by the training spatial scale. For the application that is described herein, learning at the state scale is a good trade-off: it is specific to local conditions while keeping enough data for the calibration.
