【摘要】 China’s accelerated urbanization has dramatically reshaped its urban landscape, resulting in distinct regional land-use patterns. Understanding long-term land use dynamics requires consistent, highresolution historical land use maps; however, existing datasets are limited by the constraints of spatial resolution, the scarcity of historical land use labels, and the necessity of consistent geographic units.This study proposes a cross-temporal, cross-resolution land use mapping framework to address key challenges in multi-decadal land use classification. First, super-resolution reconstruction is applied to harmonize spatial and spectral characteristics across different satellite sensors, harmonizing Landsat and Sentinel-2 data into a consistent 10 m spatial framework. Second, a semi-supervised learning framework is designed to mitigate the scarcity of historical land use labels by integrating a land use semantic memory bank, which dynamically refines pseudo-labels and enhances feature alignment. Finally, a roadconstrained multi-temporal segmentation approach is introduced to enforce spatial coherence, reducing classification inconsistencies across time-series maps. Compared to existing methods, this framework extends the temporal scope of urban land use mapping to 40 years with 5 years interval, achieving an overall accuracy ranging from 77.31% to 89.72%. Our analysis reveals a three-phase trajectory of urban land evolution: farmland-to-residential expansion in the early period(1984–2000), synchronized growth of residential, industrial, and transport land during accelerated urbanization(2000–2015), and a recent phase(2015–2024) of urban renewal and ecological restoration involving functional reconfiguration and green-oriented transitions. These findings demonstrate the shift from extensive land expansion to quality-focused restructuring, providing a critical foundation for sustainable urban planning and land governance.更多还原

【Abstract】 China’s accelerated urbanization has dramatically reshaped its urban landscape, resulting in distinct regional land-use patterns. Understanding long-term land use dynamics requires consistent, highresolution historical land use maps; however, existing datasets are limited by the constraints of spatial resolution, the scarcity of historical land use labels, and the necessity of consistent geographic units.This study proposes a cross-temporal, cross-resolution land use mapping framework to address key challenges in multi-decadal land use classification. First, super-resolution reconstruction is applied to harmonize spatial and spectral characteristics across different satellite sensors, harmonizing Landsat and Sentinel-2 data into a consistent 10 m spatial framework. Second, a semi-supervised learning framework is designed to mitigate the scarcity of historical land use labels by integrating a land use semantic memory bank, which dynamically refines pseudo-labels and enhances feature alignment. Finally, a roadconstrained multi-temporal segmentation approach is introduced to enforce spatial coherence, reducing classification inconsistencies across time-series maps. Compared to existing methods, this framework extends the temporal scope of urban land use mapping to 40 years with 5 years interval, achieving an overall accuracy ranging from 77.31% to 89.72%. Our analysis reveals a three-phase trajectory of urban land evolution: farmland-to-residential expansion in the early period(1984–2000), synchronized growth of residential, industrial, and transport land during accelerated urbanization(2000–2015), and a recent phase(2015–2024) of urban renewal and ecological restoration involving functional reconfiguration and green-oriented transitions. These findings demonstrate the shift from extensive land expansion to quality-focused restructuring, providing a critical foundation for sustainable urban planning and land governance.更多还原