Leveraging spatial structure with CapsuleNet for identification of the land use classes

Urban land use classes of complex nature are marked by the presence of multiple land covers and/or objects in the specific spatial order. The spatial configuration of the constituent parts of the land use class is generally unique. To the extent that the specific spatial configuration is defining characteristic of a given land use class. These characteristics can be effectively leveraged to identify the land use class. In this research, we exploit the unique spatial structure of the constituent parts for the land use class for its detection. We use capsule network (CapsuleNet) for detecting some of the urban land use classes such as parking lot and golf courses. CapsuleNets use a group of neurons (called capsules) in a convolutional layer to detect a specific image primitive. Each subsequent layer detects higher order primitives, and its relationship with the lower level primitives. Thus, multiple such layers build a hierarchy of parts to learn the whole object, in this case the land use class. We conducted multiple experiments for detecting parking lots and golf courses in a collection of urban images. We used NWPU-RESISC45 dataset for conducting our experiments. Furthermore, we compared the results of CapsuleNet based architecture with standard architecture such as VGG16, which do not consider the spatial structure of the features. Our initial experiments suggest improvement in accuracy in classification of the land use classes such as parking lot and golf courses.