Bal-DC LULC Atlas

Toward a Foundation Model for Ultra-High-Resolution Urban Land-Cover Classification

Junhao Wu1, Aboagye-Ntow Stephen1, Chuyuan Wang1, Michael McGuire1, Wei Yu1, Jianwu Wang2, Xin Huang1

1 Towson University    2 University of Maryland, Baltimore County

Abstract

Ultra-high-resolution (UHR) remote sensing imagery provides the spatial detail needed for fine-grained urban land-cover understanding in complex urban environments. However, existing methods are still mostly based on coarser-resolution settings or focused on downstream model adaptation for UHR imagery, without representation learning strategies designed for the unique characteristics of UHR data. In this paper, we introduce Bal-DC LULC Atlas, a framework for UHR urban land-cover understanding built on large-scale NAIP imagery. Bal-DC LULC Atlas consists of two key components: (1) UrbanMIM, a masked image modeling pretraining strategy designed for UHR urban data to capture multispectral complementarity and spatial heterogeneity in UHR imagery; and (2) the Bal-DC LULC Benchmark, a large-scale annotated benchmark for UHR urban data covering Baltimore and Washington, D.C., with 5 billion labeled pixels for systematic evaluation in complex urban environments. Extensive experiments on this benchmark show that the proposed approach learns representations that capture fine-grained and compositional spatial structures, leading to improved performance on urban land-cover understanding tasks. These results highlight the importance of UHR-specific representation learning for urban land-cover understanding.

Framework

Bal-DC LULC Atlas pipeline: NAIP RGB–NIR imagery, UrbanMIM pretraining with spectral-aware and heterogeneity-aware masking, fine-tuning on the Bal-DC benchmark, and large-scale urban LULC mapping

Bal-DC LULC Atlas connects UHR-specific masked image modeling, benchmark fine-tuning, and large-scale urban land-cover mapping.

Bal-DC LULC Benchmark

The benchmark covers diverse urban environments from sparse peri-urban landscapes to dense urban cores, with six land-cover classes annotated on 0.3 m RGB–NIR NAIP imagery.

Bal-DC benchmark examples: RGB and ground-truth patches for Baltimore County, Baltimore City, Baltimore City Core, and Washington, D.C.

0.3 m NAIP RGB–NIR imagery · Baltimore–Washington D.C. · 5B labeled pixels · 6 land-cover classes

Interactive Atlas

  • Building
  • Impervious Surface
  • Tree Canopy
  • Herbaceous Vegetation
  • Open Water
  • Bare Surface

Zoom and pan to explore NAIP RGB imagery and UrbanMIM land-cover predictions across the Baltimore region.