Gaofen-2 satellite image-based characterization of urban villages using multiple convolutional neural networks.

Remote sensing has proven to be an invaluable and effective tool for mapping urban areas. However, further efforts are required to fully utilize remote sensing data in mapping the Chinese urban villages (Chengzhongcun), which are characterized by high population density, irregular shapes and a limited view of the sky. In this study, we integrated deep neural network architectures with GaoFen-2 satellite images at a spatial resolution of 1 metre to investigate the feasibility of using FCN, Unet and ResUnet models for monitoring two categories of urban villages. The results indicate that all three deep learning algorithms demonstrate superiority in identifying and classifying urban villages, surpassing outlier identification. The urban village patches in the study areas adhere to the Zipf's law rule. Particularly, the ResUnet model outperforms FCN and Unet, achieving an overall classification accuracy of over 91% for urban villages. This study suggests that ResUnet is superior to FCN and Unet as it effectively avoids fragmentation and overfitting of individual irregular buildings within urban areas, yielding faster and more accurate results. Overall, this study underscores the significance of the Gaofen-2 dataset as an influential data source for analysing intra-urban structures. Additionally, the implementation of the ResUnet learning algorithm proves to be exceptionally valuable in extracting information across various scales and accurately capturing the intricate nature of irregular urban village patches, even when using satellite images with metre-scale resolution. [ABSTRACT FROM AUTHOR]