HiCity-LC: City-Scale Submeter Land Cover Dataset
- HiCity-LC is a high-quality annotated dataset featuring nearly 14 billion labeled pixels for submeter-resolution land cover mapping across diverse urban and rural areas.
- It employs the Mask Clustering-based Annotation Engine (MCAE) to dramatically reduce manual annotation labor while achieving classification accuracies above 85% in large-scale applications.
- The dataset covers five major Chinese cities with both dense and sparse annotations, facilitating robust transfer learning and high-fidelity mapping of fine spatial details.
HiCity-LC is a high-quality annotated dataset of about 14 billion labeled pixels for large-scale submeter land cover mapping, built with the Mask Clustering-based Annotation Engine (MCAE). It supports the generation of city-scale land cover maps across five major Chinese cities with classification accuracies above 85%, and is described as the first publicly available submeter resolution city-level land cover benchmark. Its design addresses a central bottleneck in submeter remote sensing: existing labels are typically derived from pre-existing products or manual annotation, which are often unreliable or prohibitively expensive for imagery with rich visual detail and massive data volumes (Chen et al., 29 Sep 2025).
1. Dataset scope and intended role
HiCity-LC was introduced for large-scale land cover mapping from submeter imagery, where annotation cost and label quality are major limiting factors. The dataset is explicitly tied to MCAE, which treats semantically consistent mask groups as the minimal annotating units to enable efficient, simultaneous annotation of multiple instances. In the reported formulation, the dataset is both a benchmark and a demonstration that large-scale, submeter resolution mapping can be made practical without relying primarily on pre-existing products or exhaustive object-by-object manual delineation (Chen et al., 29 Sep 2025).
The dataset is organized around city-scale mapping rather than patch collections restricted to small study areas. A common misconception is that such benchmarks mainly target dense urban cores; HiCity-LC instead covers full administrative boundaries, thus spanning urban, suburban, rural, and natural landscapes. This broader spatial scope is integral to its stated role in supporting robust transfer to unseen regions, official cropland area extraction, and city-level land cover statistics.
2. Spatial extent and data composition
HiCity-LC contains 55,887 image-label pairs, with ≈ 14 billion labeled pixels in total. Of these, 5,116 images (≈3.74 billion pixels) are densely annotated and 50,771 images (≈10.27 billion pixels) are sparsely annotated. Each image is 1024 × 1024 pixels (at 0.6 m resolution), and masks are generated from images at both 0.3 m and 0.6 m to enhance annotation. The total data volume is ≈4,432 GB (over 4 TB), and the total spatial coverage is ≈54,078 km² (Chen et al., 29 Sep 2025).
The five covered cities are Beijing, Tianjin, Shanghai, Guangzhou, and Chengdu. The per-city areas and data volumes reported for the benchmark are as follows:
| City | Area (km²) | Data Volume (GB) |
|---|---|---|
| Shanghai | 6,341 | 647 |
| Chengdu | 12,132 | 996 |
| Guangzhou | 7,434 | 607 |
| Beijing | 16,411 | 1,287 |
| Tianjin | 11,760 | 895 |
The paper also reports full image dimensions for each city-scale mosaic: Shanghai 518,656 × 446,464, Chengdu 710,144 × 502,016, Guangzhou 410,880 × 528,896, Beijing 780,032 × 603,392, and Tianjin 506,368 × 632,832. Because the coverage includes entire administrative regions rather than only built-up centers, the benchmark captures substantial intra-city heterogeneity.
3. Label system and annotation regimes
HiCity-LC adopts 8 land cover classes from OpenEarthMap's system: Bareland, Rangeland, Developed Space, Road, Tree, Water, Agricultural Land, and Building (Chen et al., 29 Sep 2025).
The annotation design combines dense and sparse regimes. Dense annotations are used for curated evaluation tiles and refinement, while sparse annotations provide large-scale supervision over much broader areas. The paper states that sparse labels are highly representative and adequate for robust large-scale model training, and further notes that sparse annotations alone already yield near-saturated performance, with Dense+Sparse providing only minor improvement. This is important for interpreting the dataset: HiCity-LC is not only a fully dense benchmark in the conventional sense, but also a study in how partial but spatially representative annotation can support city-scale semantic segmentation.
The label geometry is intended to preserve fine structures. The reported comparison emphasizes finer small-object masks and boundaries, including pathways and field ridges, that are rarely present in previous data. In this sense, HiCity-LC is not merely large; it is also constructed to retain the high-frequency spatial detail that submeter imagery makes visible.
4. MCAE annotation engine
MCAE is the annotation pipeline used to construct HiCity-LC. Its motivation is to drastically reduce manual labor and cost for submeter-resolution annotation at scale, and its design is explicitly grounded in the spatial autocorrelation principle: objects of the same class tend to co-occur with similar visual features in local neighborhoods (Chen et al., 29 Sep 2025).
The reported workflow has five steps. First, Multi-scale Mask Generation uses SAM (Segment Anything Model) at 0.3 m and 0.6 m, combining outputs for fine and coarse object boundaries. Second, Mask-level Feature Learning uses self-supervised learning to pull together features of similar objects and separate different ones, via overlapping crop contrastive strategies with a Swin Transformer backbone. Third, Hierarchical Mask Clustering groups object masks spatially and semantically using DBSCAN, with two-scale agglomeration intended to prioritize cluster purity and completeness. Fourth, Cluster-level Annotation assigns labels to pure mask clusters in one step, so that a single annotation can cover dozens of objects. Fifth, Iterative Test Set Curation samples densely annotated test tiles based on spatial stratification, then corrects and refines predictions through multiple rounds.
The efficiency claims are concrete. MCAE reduces annotation workload by 1–2 orders of magnitude over classic pixel- or mask-based manual approaches, and the cost per object is 1/67th of the mask-based baseline. In an initial experiment across test areas, the paper reports 539,512 masks in 8,057 clusters, corresponding to an average of 67 masks per cluster. Additional locality-specific examples are also given: Beijing (Daxing District): 1,036 km², 681 clusters, 23,525 masks, and Shanghai (Pudong): 1,210 km², 936 clusters, 85,646 masks.
5. Quantitative performance and evaluation
The benchmark reports Overall Accuracy (OA) above 85% in all major test areas, with values up to 92–93% in some cities. The paper also reports Mean Intersection over Union (mIoU), F1 Score, and User’s Accuracy (UA), and states that these metrics improve over baselines (Chen et al., 29 Sep 2025).
The metric definitions given in the paper are:
For the five cities, the reported MCAE OA (%) values are Beijing 92.38, Tianjin 87.80, Shanghai 85.12, Chengdu 91.93, and Guangzhou 89.23. The paper also gives city-specific mIoU examples, including Beijing: 72.06% and Shanghai: 62.74%. These results are used to support the claim that HiCity-LC enables city-scale maps with high classification accuracy despite the scale, heterogeneity, and fine-grained detail of submeter imagery.
A recurring point in the evaluation is that sparse supervision is not merely a compromise. The reported near-saturated performance with sparse annotations indicates that label placement and spatial representativeness are central variables in large-area remote-sensing annotation, not only the proportion of densely labeled pixels.
6. Relation to previous datasets, applications, and access
HiCity-LC is positioned against several pre-existing land cover datasets. The paper compares it with GID/Five-Billion-Pixels (4 m, 150 images, 50,000 km², 5.0 billion pixels), LoveDA (0.3 m, 5,987 images, 536 km², 6.0 billion pixels), and OpenEarthMap (0.25–0.5 m, 5,000 images, 799 km², 4.9 billion pixels). In this comparison, HiCity-LC is listed as 0.3–0.6 m, 55,887 images, 54,078 km², and 14.0 billion pixels, and is characterized as the largest, most detailed, city-scale, submeter dataset in that table (Chen et al., 29 Sep 2025).
The stated innovations include being the first city-scale, submeter-resolution land cover dataset for five major cities, having ≈3× more labeled pixels and ≈10× more images than OpenEarthMap, and providing consistent, scale-independent annotation through MCAE. The comparison with SinoLC-1 is also explicit: HiCity-LC is reported to exceed SinoLC-1 in OA improvement (e.g., 92% vs. 62% in Beijing), with much better fine-structure recovery, less labeling noise/propagation, and demonstrably higher cropland area estimation consistency with official statistics.
Its practical utility is presented in three directions. First, HiCity-LC is intended for training foundation models, since its size, resolution, and label quality make it suitable for pixel-level training of modern deep learning models. Second, it supports city-level mapping, including precise dense land cover maps and cropland statistics extraction. Third, it supports generalization to unseen regions, with transfer evaluated on Wuhan, alongside broader use cases in urban/rural/eco system mapping.
The dataset is publicly accessible at https://github.com/chenhaocs/MCAE. The paper excerpt states that it is publicly available but does not specify the license or detailed usage restrictions in the excerpt; the repository is the stated source for up-to-date licensing terms.