[Paper Review] LandCover.ai: Dataset for Automatic Mapping of Buildings, Woodlands and Water from Aerial Imagery.

Adrian Boguszewski, Dominik Batorski+3 more|arXiv (Cornell University)|May 5, 2020

Remote-Sensing Image Classification31 references10 citations

TL;DR

LandCover.ai introduces a publicly available, high-resolution aerial imagery dataset for semantic segmentation of buildings, woodlands, and water across 216.27 sq. km of rural Poland, with 50 cm and 25 cm per pixel resolution and fine manual annotations. The dataset enables state-of-the-art performance, achieving 90.18% mean intersection over union on a benchmark, demonstrating that cost-efficient RGB-only data can enable accurate automatic land cover mapping.

ABSTRACT

Monitoring of land cover and land use is crucial in natural resources management. Automatic visual mapping can carry enormous economic value for agriculture, forestry, or public administration. Satellite or aerial images combined with computer vision and deep learning enable the precise assessment and can significantly speed up the process of change detection. Aerial imagery usually provides images with much higher pixel resolution than satellite data allowing more detailed mapping. However, there is still a lack of aerial datasets that were made for the segmentation, covering rural area with resolution of tens centimeters per pixel, manual fine labels and highly publicly important environmental instances like buildings, woods or water. Here we introduce this http URL (Land Cover from Aerial Imagery) dataset for semantic segmentation. We collected images of 216.27 sq. km rural areas across Poland, a country in Central Europe, 39.51 sq. km with resolution 50 cm per pixel and 176.76 sq. km with resolution 25 cm per pixel and manually fine annotated three following classes of objects: buildings, woodlands, and water. Additionally, we report simple benchmark results, achieving 90.18% of mean intersection over union on the test set. It proves that the automatic mapping of land cover is possible with relatively small, cost efficient, RGB only dataset. The dataset is publicly available at this http URL

Motivation & Objective

• To address the lack of publicly available, high-resolution aerial datasets with fine annotations for semantic segmentation in rural areas.
• To support automatic land cover mapping for applications in agriculture, forestry, and public administration.
• To provide a benchmark dataset with centimeter-level resolution and precise labels for critical environmental features like buildings, woodlands, and water.
• To demonstrate that accurate semantic segmentation is achievable using only RGB aerial imagery without expensive multispectral data.

Proposed method

• Collection of 216.27 sq. km of aerial imagery from rural Poland, including 39.51 sq. km at 50 cm per pixel and 176.76 sq. km at 25 cm per pixel.
• Manual fine annotation of three semantic classes: buildings, woodlands, and water, ensuring high labeling accuracy.
• Use of standard deep learning models for semantic segmentation, trained exclusively on RGB-only input.
• Evaluation using mean intersection over union (mIoU) on a held-out test set to benchmark performance.
• Public release of the dataset at a dedicated URL to enable reproducibility and community use.
• Application of standard computer vision pipelines to assess change detection and segmentation accuracy.

Experimental results

Research questions

• RQ1 Can a relatively small, publicly available, RGB-only aerial dataset achieve high performance in semantic segmentation of key land cover types?
• RQ2 How does the inclusion of high-resolution (25 cm) imagery impact segmentation accuracy compared to lower-resolution (50 cm) data?
• RQ3 To what extent can fine-grained manual annotations improve model generalization in rural land cover mapping?
• RQ4 Can automatic mapping of buildings, woodlands, and water be achieved with cost-efficient data collection and processing pipelines?

Key findings

• The LandCover.ai dataset achieves a mean intersection over union (mIoU) score of 90.18% on the test set, indicating high segmentation accuracy.
• The dataset includes 216.27 sq. km of rural Polish terrain with 39.51 sq. km at 50 cm per pixel and 176.76 sq. km at 25 cm per pixel.
• Fine manual annotations for three critical land cover classes—buildings, woodlands, and water—were successfully created at scale.
• The results demonstrate that high-accuracy semantic segmentation is feasible using only RGB aerial imagery without multispectral or LiDAR data.
• The dataset is publicly available, enabling reproducible research and benchmarking in land cover mapping.
• The benchmark results confirm that cost-efficient, high-resolution RGB data can support precise and scalable land cover assessment.