Source-linked AI summary
LandCover.ai: Dataset for Automatic Mapping of Buildings, Woodlands, Water and Roads from Aerial Imagery
Adrian Boguszewski, Dominik Batorski, Natalia Ziemba-Jankowska, Tomasz Dziedzic, Anna Zambrzycka
TL;DR
Fine-annotated aerial datasets covering rural areas and the four land-cover classes are lacking, despite the value of land-cover monitoring and the time and error associated with manual change detection. LandCover.ai introduces a manually labeled Polish aerial dataset and reports an 85.56% test-set mIoU benchmark. The authors conclude it supports automatic mapping and can help improve the efficiency and accuracy of identifying land-cover changes.
Problem
Fine-annotated aerial datasets for rural areas and simultaneous segmentation of buildings, woodlands, water, and roads are lacking, while manual change detection is time-consuming and error-prone.
Source: §1 · p. 2Method
Results
LandCover.ai fills an open-aerial-dataset gap, and the augmented baseline reaches 85.56% mIoU on the test set.
Source: §5 · p. 6, §4.4 · p. 6Takeaways & Limitations
The authors conclude the dataset can support neural-network mapping and help improve the efficiency and accuracy of identifying land-cover changes.
Source: §5 · p. 6Roads and buildings are challenging to segment because they are often narrow or small, have fewer easily classified inner pixels, and may be obscured by trees.
Source: §4.4 · p. 6
Abstract
from arXiv · showhide
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 precise assessment and can significantly speed up 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 made for the segmentation, covering rural areas with a resolution of tens centimeters per pixel, manual fine labels, and highly publicly important environmental instances like buildings, woods, water, or roads. Here we introduce LandCover.ai (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 four following classes of objects: buildings, woodlands, water, and roads. Additionally, we report simple benchmark results, achieving 85.56% of mean intersection over union on the test set. It proves that the automatic mapping of land cover is possible with a relatively small, cost-efficient, RGB-only dataset. The dataset is publicly available at https://landcover.ai.linuxpolska.com/
1. Introduction
Land-cover information supports resource management and monitoring, but the manual change-detection workflow is slow and error-prone. LandCover.ai addresses the shortage of fine-annotated aerial data covering four important classes.
1. Introduction
- Land-cover information supports resource management and can inform studies of environmental change, forest dynamics, degradation, and biodiversity monitoring.Parcel imagery also has economic value for agriculture and public administration, where land use affects subsidies and taxes.Source: §1 · pp. 1–2
- Manual change detection can take months nationwide; the authors report average error above 30% over multiple years for an area exceeding 312,000 km2.They describe this work as expensive and argue for an efficient automatic tool.Source: §1 · p. 2
- Deep learning can support scalable aerial-image change detection, but usually requires large ground-truth datasets, while fine-annotated natural datasets remain scarce.The cited related work includes a weakly labeled approach for detecting buildings and roads from RGB orthophotos.Source: §1 · p. 2
- LandCover.ai provides manually annotated aerial imagery for buildings, woodlands, water, and roads across 216.27 km2 in Poland, with a baseline benchmark.The coverage includes 39.51 km2 at 50 cm per pixel and 176.76 km2 at 25 cm per pixel.Source: §1 · p. 2
2. Related works
Prior aerial and satellite datasets cover varied tasks and land-cover categories, but the related-work review identifies a gap in rural, fine-labeled datasets spanning the four classes targeted by LandCover.ai.
2. Related works
- Many aerial or satellite datasets target whole-image classification, while DOTA and iSAID address multi-class detection and instance segmentation rather than the public-agency land-cover needs discussed here.The review notes that common-object and street-view datasets also differ from aerial and satellite imagery in perspective and class needs.Source: §2 · p. 2
- Several aerial datasets focus on buildings or vegetation alone, limiting their coverage of the combined classes needed for land-cover change detection.The review lists Massachusetts, Inria, AIRS, Agriculture-Vision, and Tree Cover among these datasets.Source: §2 · p. 2
- ISPRS Vaihingen and Potsdam offer manual labels but cover fewer than 5 km2 and omit water; Chesapeake Bay is much larger but has coarse automatic labels and no building class.The review describes Chesapeake Bay at about 160,000 km2 and 1 m resolution.Source: §2 · pp. 2–3
- Some UAV image and video datasets are designed for event recognition, so the review considers them inapplicable to this segmentation task.Examples include ERA, UCLA Aerial Eve, and Okutama-Action.Source: §2 · p. 3
3. The dataset
LandCover.ai is a manually labeled, RGB-only aerial dataset designed to support semantic segmentation of four land-cover classes in rural Polish imagery. It combines broad geographic and image-condition variation with fine annotations, addressing a gap in related datasets.
3.1. Data acquisition
- The dataset comprises 41 manually selected orthophoto tiles from counties in all regions, totaling 216.27 km2: 176.76 km2 across 33 images at 25 cm resolution and 39.51 km2 across 8 images at 50 cm resolution.Selected images vary in region, season, time of day, weather, and lighting conditions.Source: §3.1 · p. 3, Figure 2 · p. 3
3.2. Land cover characteristics
- The selected Polish areas span varied geography, from northern and central lowlands to southern highlands and mountains, with agricultural land and forests prominent.The passage reports agricultural areas at 60% and forests at 29.6% of the landscape.Source: §3.2 · p. 3
3.3. Classes
- The four classes are buildings, woodlands, water, and roads, defined to serve public-administration needs.Buildings include roofs and visible walls; woodlands exclude single trees and orchards, water excludes ditches and dry riverbeds, and roads include parking, unpaved roads, and rail tracks.Source: §3.3 · p. 4
3.4. Annotations
- Annotations were made manually on image tiles, reviewed by a second person, and converted into image-level segmentation masks.Narrow roads were traced as polylines and converted to fixed-thickness polygons; the dataset includes 12,280 buildings and 72.02 km2 of woodland.Source: §3.4 · p. 4
3.5. Comparison to related datasets
- The comparison identifies no existing dataset combining these four classes, rural coverage, tens-of-centimeters resolution, and fine manual labels.Related datasets were described as single-class, urban-only, lower-resolution, or automatically and coarsely labeled.Source: §3.5 · p. 4
4. Experiments
Experiments benchmarked DeepLabv3+ on LandCover.ai, testing data augmentation and model modifications; augmentation raised test-set mIoU to 85.56%.
4. Experiments
- DeepLabv3+ with a modified Xception71 and Dense Prediction Cell backbone served as the baseline for testing model modifications and augmentation.The experiments aimed to assess whether these changes could improve results and by how much.Source: §4 · p. 4
4.1. Data preparation
- The authors split 41 images into 512×512 tiles, assigning 15% each to test and validation and 70% to training.This produced 1,602 test tiles, 1,602 validation tiles, and 7,470 training tiles.Source: §4.1 · p. 5
4.2. Data augmentation
- Offline augmentation added nine copies per training tile using color, image-quality, geometric, cropping, and padding transformations.The resulting training set contained 74,700 samples; example augmentations simulate variation in seasons, lighting, haziness, and object sizes.Source: §4.2 · pp. 5–6, Figure 6 · p. 6
4.3. Training
- Training used Cityscapes-pretrained Xception71 DPC, decoder output stride 4, and class-weighted loss to counter imbalanced class areas.Training stopped when validation mIoU showed no significant gain; encoder output stride, batch size, and atrous rates were varied.Source: §4.3 · p. 6
4.4. Results
- Roads and buildings are the hardest segmentation classes because narrow roads and small buildings have fewer easily classified inner pixels, making imprecise edges cause greater error; trees sometimes obscure them.Table 2 reports test-set results using IoU metrics. Figure 7 shows close-ups of images, labels, and result segmentation.Source: §4.4 · p. 6
- 85.56% mIoU was reached after augmentation improved metrics by 1.47%; the smaller output stride gave 84.09%.The baseline DeepLabv3+ model reached 81.81% mIoU on the entire test set.Source: §4.4 · p. 6
5. Conclusions
LandCover.ai provides manually annotated, RGB-only aerial imagery for four land-cover classes across rural Poland. The authors report that neural networks can use the dataset for automatic mapping, and make it publicly available.
5. Conclusions
- LandCover.ai is an RGB-only dataset of high-resolution aerial imagery, manually annotated for buildings, woodlands, water, and roads.The images cover varied rural areas and optical conditions across vegetation seasons.Source: §5 · p. 6
- The authors report that the dataset can support neural-network tools for automatic mapping and more efficient and accurate identification of land-cover changes.They identify administration, agriculture, forestry, and water resource management as potential application areas.Source: §5 · p. 6
- The authors plan to add classes such as fields and ditches and divide water into categories such as lakes, rivers, ponds, and pools.Source: §5 · p. 6
- The dataset is publicly available to encourage its future development and use.Source: §5 · p. 6