DeSARD: A Dataset for Aerial Human Detection in Desert Search and Rescue

Unmanned aerial vehicles (UAVs) can support search and rescue (SAR) in desert environments, but progress is limited by the lack of public datasets for aerial human detection in arid terrain. We present DeSARD, a real and synthetic dataset for UAV-based human detection in desert SAR scenarios. The real subset contains 7,056 UAV images collected in desert terrain with 3,420 human bounding boxes and altitude metadata spanning 20--95 m. To complement the UAV imagery, DeSARD includes a procedurally generated synthetic subset created in Blender with randomized terrain, vegetation, lighting, human pose, camera height, and camera-quality post-processing, together with automatically generated annotations and metadata. We release 5,000 synthetic images together with their annotations the procedural generation pipeline used to create them. Technical validation using image classification and object detection shows that the real subset supports both image-level and object-level human-detection workflows. Synthetic augmentation produced model-dependent benefits by improving lightweight classifiers, with benefits dependent on the balance between real and synthetic training images. Overall, DeSARD provides a public resource for desert SAR perception, altitude-aware aerial analysis, and sim-to-real augmentation studies.