Influence of Geographic Distance on CNN Generalization for Satellite Image Classification

The remote sensing research community is grappling with methods to produce training data that are sufficiently representative of large areas to which they want to scale up their machine learning models for image classification. Effective generalization, will allow land cover classification models trained in data-rich regions to be applied to data-poor regions, with minimal increases in error. This study investigated cross-location generalization through model transfer of convolutional neural networks (CNN), in a series of experiments spread across eight counties within the Chesapeake Bay Catchment. The model transfer was effective (> 80% accuracy), even with as little training as 80/class, across distances up to 600 km. Classification accuracy and to a lesser extent, image similarity in CNN feature space, decreased with geographic distance, but are not the overriding factors governing model transfer.