Version: 1.0 License: CC BY-NC 4.0 (with ODbL terms for derivative geospatial data)
Currently being updated to expand to all countries, sift through some noise, add a more scalable annotation process, and improving VLM (overall), but also to not prioritize geographic name over visual features. Please hold tight for the update and resulting benchmarked results :).
RoofNet is the largest and most geographically diverse open-access dataset for global roof material classification. It consists of high-resolution Earth Observation (EO) image tiles paired with structured metadata and curated textual prompts describing roofing characteristics across 14 material classes and an "Unknown" class. The dataset is designed to support hazard preparedness, resilience planning, and post-disaster supply-chain analysis and will have benchmarked results shortly.
RoofNet includes:
-51,503 EO image tiles spanning 184 urban regions across 112 countries
-14 roof material classes (e.g., Asphalt Tiles, Clay Tiles, Metal Sheets, Thatch, etc.)
-A multimodal CSV file with rich per-sample metadata
-RemoteCLIP ViT-L/14 models fine-tuned for roof classification
-Training and evaluation code for reproducible fine-tuning and VLM experimentation.
RoofNet includes 14 roofing material classes grouped into 5 categories:
- Natural/Traditional: Thatch, Green Vegetative
- Stone/Ceramic Tiles: Stone Slates, Clay Tiles
- Asphalt/Concrete/Wood Tiles: Asphalt Tiles, Concrete Tiles, Wood Tiles
- Sheet-Based: Metal Sheet Materials, Polycarbonate Sheet Materials, Glass Sheet Materials
- Synthetic/Amorphous: Amorphous Asphalt, Amorphous Concrete, Amorphous Membrane, Amorphous Fabric
Fine-Tuned Models The models/ folder includes a fine-tuned version of RemoteCLIP ViT-L/14, adapted using 6,000 manually annotated samples, class re-balancing, and with prompts incorporating geographic and material cues. See the notebooks folder for reproducible training and evaluation pipelines.