Automated High Resolution Land Cover Generation from WorldView Multispectral Imagery
To date, many efforts in generating reliable land cover classification from multispectral imagery (MSI) have focused primarily on 30 meter resolution Landsat imagery. At this resolution, it is possible to differentiate large scale man-made and natural features including cities and forests, however, higher resolution objects are difficult to extract. We present a neural network trained on spectral and spatial information in five distinct world regions using WorldView-2 8-band MSI. Our network is capable of classifying objects such as individual trees, roofs, irrigation ponds, etc. that have resolutions higher than 2 meters. Our process co-registers and aggregates multiple overlapping MSI classification images to improve overall classification. Aggregated classification images are registered to OSM road vectors to provide reliable high resolution classifications that can be used in a variety of applications such as automated object detection (i.e. new or missing road/building objects in OSM), change detection, and urban planning to name a few.