Multi-temporal high-resolution urban land-use mapping and change analysis based on a deep geospatial-temporal adaptation network

Automated mapping and change analysis of urban land use are crucial tasks for examining the patterns of urban development and effectively directing the sustainable management of urban land resources. High-resolution (HR) remote sensing imagery offers abundant spatial details and clear urban structures. However, the existing change detection methods require high-quality paired samples and are based on the assumption that the training and test data are independent and identically distributed, and thus lack the flexibility to generalize the trained model to new temporal images. In response to the challenge, a multi-temporal urban scene classification and change detection (MtUS-CCD) framework is proposed to realize urban land-use mapping and change analysis, with the real geographic boundaries provided by OpenStreetMap (OSM) road networks. The key model of the proposed MtUS-CCD framework is the deep geospatial-temporal Adaptation Network based on partial selftraIning and geospatial-Temporal Alignment (ANITA). The ANITA model employs a geospatial-temporal alignment (GTA) strategy to align the geographical locations of multi-temporal images, acquiring deep features that are invariant to temporal domain shifts. Label migration and self-training classification (STC) are also performed to enhance the model's discriminative capacity for cross-temporal urban scene classification in images obtained from new time phases. To relieve the significant scale differences and high shape variability among urban parcels, the ANITA model leverages the area-weighted voting (AWV) strategy to achieve land-use mapping based on the multi-temporal comprehensive OSM road network data. Subsequently, post-classification comparison (PCC) enables the acquisition of the land-use change directions. The experimental results obtained on tritemporal datasets from China demonstrate that the MtUS-CCD framework shows a significant improvement in cross-temporal urban scene classification and change detection tasks conducted in different regions. Furthermore, this framework shows robust effectiveness and generalization in a large-scale application for the whole of the city of Wuhan in China. Through comparative analysis with policy planning, it is demonstrated that the urban development patterns inferred by this framework are accurate and reliable, providing strong support for the realization of sustainable development goals.