Knowledge-guided land pattern depiction for urban land use mapping: A case study of Chinese cities

Accurate urban land-use maps, which reflect the complicated land-use pattern implied in the function and distribution of land-cover types, play an important role in urban analysis. In recent years, data-driven deep learning-based land-use mapping methods have made great breakthroughs due to their strong feature extraction ability. Meanwhile, multisource geographic data, such as open street map (OSM), has been applied in land-use mapping with high spatial resolution remote sensing (HSR) imagery. Nevertheless, given the intraclass visual inconsistency and interclass label ambiguity, there are enormous challenges in OSM-based land-use pattern depiction: 1) the significant size variability of land-use parcel generated by OSM; 2) the weak interpretability of the data-driven based features; and 3) the neglect of intrinsically hierarchical and nested relationships between land-cover and land-use. In this paper, to bridge the “knowledge gap” for urban land-use mapping, a knowledge-guided land pattern depicting (KGLPD) framework is proposed. The proposed KGLPD framework mainly contains four parts. Land-use parcels with various scales are generated based on OSM. An adaptive gradient perceptive (AGP) mechanism is proposed to provide patch distribution prior knowledge for guiding the data-driven based visual feature extraction. To effectively cognize the layout of different land-cover types as the knowledge-driven information, a land pattern cognitive (LPC) model is designed to capture the inner and outer relationships (i.e., direction, distance and co-frequency) of different land-cover types. The fully sparse topic model (FSTM) is then used to extract the critical land pattern information from the data-driven and knowledge-driven information. Four typical Chinese urban cities are selected to evaluate the proposed framework. Experimental results on three cities with four regions of distinctive characteristics in different years, have achieved high classification accuracies of about 80%, with 10% improvement compared with other methods. This demonstrates the effectiveness and robustness of the proposed knowledge-guided urban land use mapping framework. Experimental results on the whole city of Shenzhen in China imply that the proposed framework perform well with small training samples. The results on different cities validate the generalizability and transferability of KGLPD. The typical land-use maps and the corresponding land-cover maps help understanding the relationship between them.