Geographically distributed data naturally varies from one location to another due to different environmental conditions between regions. This creates a representation or covariate shift in input variables between training and testing data when we apply a model to a different location. Theoretically, we expect this covariate shift to have a detrimental impact on model performance. However, this negative impact is hard to estimate beforehand merely from the input data, and trained models may perform surprisingly well even under distribution shifts.This paper investigates how different covariate shift strategies impact the model performance on geospatial vegetation forecasting. In our experiments, we demonstrate that the model accurately predicts in locations far from the training samples in space by leveraging the similar ecological behavior of vegetation under comparable environmental conditions. We close with an extensive summary that outlines our findings and provides an outlook on discussion points that we hope to discuss in depth at the workshop.