A Framework for Land Use Scenes Classification Based on Landscape Photos

Space-earth integrated stereoscopic mapping promotes the progress of earth observation technologies. The method which combined remote sensing images with zenith perspectives and ground-level landscape photos with slanted viewing angles improves the efficiency and accuracy of land surveys. Recently, numerous efforts have been devoted to combining deep learning and remote sensing images for the classification of land use scenes. However, improvement of classification accuracy has been limited because of the lack of sectional representation. Landscape photos can describe the cross-sections in detail. For this reason, this study constructed a land-use semantic photo dataset (LSPD) and proposed a land-use classification framework for photos (LUCFP) based on Inception-v4. LSPD was constructed through semantic planning, scene segmentation, supervised iteration transfer learning, and augmentation of photos. LSPD has 1.4 million photos collected from seven geographic regions of China, and covers 13 land-use categories and 44 semantic categories. LUCFP adapts scene segmentation based on depth of field, multisemantic block labeling, and weighting of semantic joint spatial ranges to determine the land use category. To validate LUCFP, nine semantic samples (9×3×2000 photos) were chosen from LSPD, obtaining an overall accuracy of 97.64%. The best photo cropping method was masking, which crops the boundary of the scene labeled by the photo, leading to an accuracy of 90.32%. The optimal pixel size that balances speed and accuracy is 675×675, with speed reaching 30 photos per second with an average accuracy of 93.80%. LUCFP has been successfully applied to the automatic verification of land surveys in China.