Updating cadastral maps using deep convolutional networks and hyperspectral imaging

Cadastral maps are maps showing the boundaries and ownership of land parcels. Surveying and updating cadastral maps is typically performed using several conventional methods such as field work and remote sensing (RS) based on aerial and satellite images. While field surveying is accurate, it is time-consuming and impossible when access to remote areas is difficult due to harsh weather conditions or other restrictions. Therefore, RS is often preferred to field survey. To update a cadastral map, land use information on how people use the landscape is required. However, land use cannot be determined using RS imaging. RS imaging can only generate land cover information detailing how much of a region is covered by forests, wetlands, impervious surfaces, and other land types. In this study, we employed hyperspectral images and a deep learning network to efficiently use RS imaging for updating a cadastral map. The classes output by the network trained on hyperspectral images were reorganized according to the cadastral categories to update the map. Through this process, it was possible to extract the areas requiring updating, and update their attribute information. The results demonstrated that hyperspectral images could be effectively classified, allowing to update the considered cadastral map.