Automatic generation of land use maps using aerial orthoimages and building floor data with a Conv-Depth Block (CDB) ResU-Net architecture

Unlike land classification maps, it is difficult to automate the generation of land use (LU) maps. The deep learning approach is a state-of-the-art methodology that can expedite the creation of LU maps. However, as the deep learning output depends on the training input, it is critical to decide upon the input that should be selected. In this study, a method for securing accurate LU information is established and used for ground truthing, using data on the number of building floors extracted from a digital topographic map and a 51 cm resolution aerial orthoimages as inputs. To this end, we developed a Conv-Depth Block (CDB) ResU-Net architecture. To verify the versatility of the proposed network, our neural network was applied to three complex metropolitan areas with different LU characteristics in Korea. The accuracy of LU maps for these cities was improved by combining convolution layers and depth-wise separable convolution as well as by including numerical building floor data. The proposed CDB ResU-Net achieved an overall accuracy of 83.7 % for the test samples. Our network exhibited an improved performance compared to Deeplab v3+, ResUnet, ResASPP-Unet, and context-based ResU-Net in classifying residential classes, which is crucial for estimating the degree of exposure in urban risk analyses.