Assessing the generalization capability of deep learning networks for aerial image classification using landscape metrics

•This method checks the validity of a trained deep learning network on a new image.•Landscape metrics (LMs) compare training images and the unseen test image.•LMs are calculated from reference data, or unsupervised clustering or segmentation.•Similarity of LMs is correlated to the accuracy of a trained network on a new image.•Results are presented for a building classification task for three African cities.