RoofSceneX: Cross‑Condition Open Dataset for Fine‑Grained Building Occupancy Classification and Rooftop Segmentation from Satellite Imagery under Occlusion and Adverse Weather

RoofSceneX RoofSceneX is a high-resolution global satellite imagery dataset developed for building rooftop segmentation and building occupancy classification. The dataset contains manually annotated rooftop masks and occupancy labels collected from nine geographically diverse cities worldwide under challenging imaging conditions including cloud cover, vegetation occlusion, shadowing, roof complexity, and dense urban patterns. Dataset Characteristics 49,283 manually annotated rooftop instances 9 cities worldwide 30 cm spatial resolution 512 × 512 image tiles Binary rooftop masks for segmentation Color-coded occupancy labels for classification Supported Tasks Building rooftop semantic segmentation Building occupancy detection Object detection Urban building taxonomy creation Included Data Sample satellite image tiles Binary rooftop masks Occupancy annotations Metadata Model scripts Trained Models This repository includes trained benchmark models developed in this study: MobileUNet+ UNet AttUNet+ ViTAttUNet+ YOLO26 family for occupancy detection Citation Please cite the work in case of use: the reference will be added after related paper gets published.