How a Million‑Scale 159‑Category Dataset and Foundation Model Set New Standards for Remote‑Sensing Object Detection

Remote‑sensing object detection underpins critical applications such as urban monitoring, traffic analysis, maritime surveillance, airport management, and emergency rescue, yet it has long suffered from limited dataset scale, fragmented category systems, and large sensor‑resolution gaps that hinder model generalization.

To address these bottlenecks, Professor Shi Zhenwei and Professor Zou Zhengxia’s team at Beihang University released LEVIRDet‑159, a million‑scale, 159‑category remote‑sensing detection dataset. The dataset contains 174,488 images (over 1.735 × 10¹¹ pixels) and 2,563,973 annotated objects, covering 30 common parent classes and fine‑grained subclasses (45 aircraft types, 13 vehicle types, 71 ship types). It aggregates imagery from satellites (SPOT, Gaofen series, Pleiades), aerial platforms, and map services (Google, Baidu, Bing), providing roughly 70 k fine‑grained labels and ensuring cross‑platform diversity.

Beyond sheer size, the authors built a unified annotation engine that standardizes bounding‑box formats, introduces a multi‑level taxonomy, applies source‑blind fine‑grained re‑annotation, and performs whole‑image consistency checks. Approximately 1.48 M boxes are newly added, and 0.85 M boxes are geometrically corrected or re‑labeled, meaning the dataset is a systematic reconstruction rather than a simple merger of existing collections.

Leveraging LEVIRDet‑159, the team designed LEVIRDetNet, a scale‑ and hierarchy‑aware foundation model for universal remote‑sensing detection. Its three core innovations are:

Online prediction of visual ground‑sampling distance (GSD) directly from image content, supplying a scale condition signal without external metadata.

GSD‑conditioned query modulation and dynamic query allocation, allowing the model to reduce redundant computation in sparse scenes while preserving detection capacity for dense small‑object scenarios.

Hierarchical‑aware detection head that jointly exploits parent‑class, child‑class, and ancestor‑path information during training, mitigating errors caused by flat classifiers in mixed‑granularity label spaces.

To evaluate true generalization, the authors adopted a strict target‑training‑free cross‑benchmark protocol: LEVIRDetNet is trained only once on LEVIRDet‑159 and then tested unchanged on nine external remote‑sensing detection benchmarks, without any fine‑tuning or use of external training images.

Under this setting, LEVIRDetNet achieves first place on all nine benchmarks, attaining an average AP of 80.56 % and surpassing the strongest fully supervised competitor by 5.02 mAP. Compared with open‑set detection and grounding models, it also shows more stable precision‑recall trade‑offs, especially at confidence thresholds common in real deployments.

The results demonstrate that LEVIRDetNet does not merely “overfit” a single dataset; it generalizes across varied category systems, spatial resolutions, and sensor platforms, fulfilling the promise of a true foundation model for remote‑sensing detection.

Looking forward, the release of the complete dataset, annotation license list, code, and pretrained weights is expected to accelerate research on universal remote‑sensing detection, fine‑grained recognition, cross‑domain interpretation, and the broader development of vision foundation models for Earth observation.