Mapping China’s cities at submeter precision

Land use and cover (LULC) information underpins studies on climate science, disaster management, food security and ecosystem protection. Advances in satellite imagery have improved resolution, but high-resolution land cover mapping still faces major obstacles.

Traditional machine learning methods often fail to capture fine-grained urban structures, while deep learning approaches require huge amounts of labeled data, which are laborious and expensive to produce. Weakly supervised foundation methods and models are promising, but struggle to be accurate and transferable to diverse urban environments. Due to these challenges, there is an urgent need to develop new strategies that reduce annotation costs while ensuring reliable submetric mapping at large scales.

A team from Wuhan University and Zhejiang Mingzhou Institute of Surveying and Mapping has developed a new solution. Their study, published in the Remote Sensing Journalintroduces the “Initial and Extended Labeling” (IEL) engine and introduces EcoVision, a submeter resolution land cover product covering 42 of China’s largest cities. By integrating high-resolution imagery, collaborative data, and deep learning models, researchers obtained an unprecedented 0.5-meter land cover dataset with 83.6% accuracy across more than 23 million validation pixels.

The IEL annotation engine works in two stages. First, it generates reliable “seed labels” through a priority-based weighted voting strategy, reconciling multiple historical land cover products with different resolutions and classification schemes. Second, it iteratively extends these labels using a semantic segmentation network, gradually refining the accuracy through repeated cycles until the model’s performance stabilizes. This hybrid approach overcomes incompatible formats, pixel misalignment, and label sparsity.

Applying IEL, the team produced EcoVision, which classifies urban areas into eight categories: buildings, roads, other impervious surfaces, grass/shrubs, trees, soil, agriculture and water. The validation included 2,385 image patches covering 23,850,000 pixels in 42 cities, providing an overall accuracy of 83.6%. Key categories such as buildings achieved an F1 score of 90%, while roads and agriculture exceeded 83%.

Compared with five leading products including Hi-ULCM and SinoLC-1, EcoVision offered higher resolution and richer details, including distinguishing impervious surfaces and extracting agricultural plots. Visual comparisons highlighted EcoVision’s ability to accurately delineate roads obscured by shadows and capture fine urban ecological mosaics. This makes it the first large-scale submeter LULC product available for China, now released publicly as an open dataset.

“EcoVision represents a milestone in urban remote sensing,” said lead author Encheng Zhang. “By eliminating the bottleneck of manual labeling, our IEL engine enables the creation of high-resolution land cover maps at previously unattainable scales. The precision and detail of EcoVision allows us to see Chinese cities as dynamic ecological systems, not just built structures. protection of ecosystems. »

The EcoVision release provides an invaluable tool for several areas. Urban planners can use it to assess the distribution of green space and infrastructure expansion, while environmental scientists can study heat islands, carbon storage and water dynamics. Policymakers can apply the dataset to land use planning, climate adaptation strategies and monitoring sustainable development goals.

The open availability of 0.5 meter land cover data in 42 major cities also creates opportunities for machine learning applications, such as urban growth modeling and disaster risk assessment. Ultimately, EcoVision demonstrates how innovative AI-driven annotation can unlock the full potential of high-resolution imagery to shape resilient and sustainable cities.