NeoDrone

NEODrone (Near-Earth Observe via Drone Dataset) is a large-scale, multi-modal, multi-category, and metadata-rich drone-based near-Earth observation dataset designed for complex low-altitude environments. It aims to support cutting-edge research in low-altitude visual perception, multi-modal fusion, and domain adaptation. Data collection was conducted between May 2021 and May 2022 in Hebei Province, China, covering five representative low-altitude scenarios: urban roads, rural farmland, water bodies, desert and Gobi regions, and forested areas. The data were captured under diverse weather conditions—including clear, cloudy, overcast, foggy, and light rain—and across multiple times of day (early morning, morning, noon, afternoon, and evening) and all four seasons (spring, summer, autumn, and winter), ensuring high environmental and scene diversity. All flight missions were operated by professional pilots holding the Civil UAV Pilot License issued by the Civil Aviation Administration of China (CAAC), using a DJI Mavic 3T UAV platform at relative flight altitudes ranging from 30 to 120 meters. The platform integrates three synchronized sensors: (1) a 48-megapixel wide-angle visible-light camera (24 mm equivalent focal length, 84° diagonal field of view), (2) a 12-megapixel telephoto visible-light camera (162 mm equivalent focal length), and (3) a long-wave infrared (LWIR, 8–14 μm) thermal imaging camera with a native resolution of 640×512 pixels (supporting 1280×1024 super-resolution mode). These sensors are mounted on a high-precision three-axis gimbal, achieving hardware-level time synchronization with a synchronization error of less than 10 ms. Raw data were recorded as 1920×1080 video at 30 frames per second (FPS), totaling approximately 1.2 TB. The data processing pipeline included adaptive keyframe extraction, a dual-stage cleaning process combining automated pre-screening with manual review, and a three-stage annotation workflow involving AI-assisted labeling, manual calibration, and expert quality inspection. The final dataset comprises 150,000 high-quality images, including 147,000 visible-light images and 3,000 infrared images, forming 3,000 strictly spatiotemporally aligned multi-modal image pairs. All images are annotated with axis-aligned bounding boxes in Pascal VOC format, covering 30 object categories (e.g., various vehicles, construction machinery, vessels, UAVs, and pedestrians), with over 1.5 million annotated object instances in total. Each image group (derived from the same video segment) is associated with a structured metadata file in .xlsx format, containing 15 key parameters: dataset type, update date, acquisition height (in meters), data source type (visible/infrared), acquisition location (latitude and longitude), terrain type, weather condition, time of day, season, azimuth angle, scene description, image format (JPG), annotation format (XML), category names, and corresponding integer labels. Metadata completeness is 100%, ensured through a hybrid approach combining automatic logging (via UAV API and image EXIF data) and manual entry. Spatial information is recorded in the WGS-84 coordinate system (e.g., 34°26′N, 108°94′E), and temporal information is precise to the acquisition date (YYYYMMDD). Image spatial resolution is uniformly 1920×1080 pixels. There is no systematic data loss; any frames removed during cleaning (e.g., due to blur, overexposure, absence of targets, severe occlusion, or sensor anomalies) were synchronously excluded along with their multi-modal counterparts and metadata, ensuring high consistency in the released dataset. Data errors primarily stem from natural environmental factors (e.g., atmospheric turbulence affecting infrared imaging) and subjective variability in manual annotation; however, these are minimized through a rigorous five-level quality control process. The dataset is organized by video segment, with each directory containing four subdirectories: Data/images/light (visible-light images), /red (infrared images), Annotations/ (XML annotation files), and metadata/ (metadata Excel tables). All files use standard, widely supported formats (JPG, XML, XLSX) and can be directly read by mainstream computer vision frameworks (e.g., MMDetection, YOLO, Detectron2) and common office software, requiring no specialized tools. Internal technical validation confirms that NEODrone demonstrates excellent performance and broad application potential in object detection, multi-modal fusion, and domain generalization tasks.