Testing Spatial Out-of-Sample Area of Influence for Grain Forecasting Models: How does out of Spatial Out-of-Sample AoI Change through the Season?

The potential for predictive models based on earth observations (EO) and survey data to assist in famine early warning and other development applications is rapidly growing. However, while the spatial-temporal extent of EO data is complete, high quality survey data is generally limited in spatial and temporal scope. The perennial question in all predictive analysis, and especially when trying to move from research to operational application in the developing world is: If we create a forecast model from region A (based on observed outcomes) can we apply the same model in region B, where we do not observe or have limited observations of those outcomes? Prior research has proposed examining the Area of Influence (AoI) based on structurally similar characteristics in the EO predictors. We expand on and evaluate this approach in the context of grain yield forecasting in Sub-Saharan Africa (SSA). Specifically, we evaluate an AoI methodology established for generating raster surfaces and apply it to vector supported grain data.  We ask the following questions: What are the key characteristics that make a forecast fit for one country work in another country? Can pooling models across multiple countries provide more accurate out-of-sample estimates than a model fit to one country or district? Does AoI change through the season? Does a model fit for in early season have the same AoI as a model fit late in the season.