Junjue-Wang/LoveDA

What it solves

LoveDA is a high-resolution remote sensing dataset designed to help AI models better identify land-cover types across different environments. It specifically addresses the challenge of domain adaptation—ensuring a model trained on one type of area (like an urban center) still works accurately when applied to a different area (like a rural region).

How it works

The project provides 5,987 high-resolution (0.3m) images from three Chinese cities (Nanjing, Changzhou, and Wuhan). These images are categorized into seven labels: background, building, road, water, barren land, forest, and agriculture. Researchers use these images to train and test semantic segmentation models, which assign a class label to every pixel in an image to map land cover.

Who it’s for

This dataset is intended for researchers and developers working on computer vision for remote sensing, specifically those focusing on semantic segmentation and unsupervised domain adaptation.

Highlights

• High Resolution: Images have a spatial resolution of 0.3 meters.
• Geographical Diversity: Focuses on the contrast between urban and rural environments.
• Complex Challenges: Includes multi-scale objects, complex backgrounds, and inconsistent class distributions to test model robustness.
• Integration: Included in popular libraries like Torchgeo and MMsegmentation.