Abstract
Satellite imagery and remote sensing provide explanatory variables atrelatively high resolutions for modeling geospatial phenomena, yet regionalsummaries are often desirable for analysis and actionable insight. In thispaper, we propose a novel method of inducing spatial aggregations as acomponent of the machine learning process, yielding regional model featureswhose construction is driven by model prediction performance rather than priorassumptions. Our results demonstrate that Genetic Programming is particularlywell suited to this type of feature construction because it can automaticallysynthesize appropriate aggregations, as well as better incorporate them intopredictive models compared to other regression methods we tested. In ourexperiments we consider a specific problem instance and real-world datasetrelevant to predicting snow properties in high-mountain Asia.