Small-object Domain-adaptive Detection in UAV Imagery: A Review, Challenges, and Prospects

Hu Zihan

School of Computer Science and Technology, Beijing Jiaotong University

Abstract：

In recent years, with the steady advance of deep learning, small-object detection in UAV imagery has matured rapidly and found wide use in smart cities, post-disaster search and rescue, criminal investigation, and agricultural management. Yet the particular imaging geometry of UAV platforms makes real-world deployment difficult. Once the viewpoint changes, the target-sensor distance grows, or weather deteriorates, deep detectors often suffer from pronounced domain shift and progressive disappearance of small-object features. Most existing studies and surveys treat small-object detection and domain adaptation as separate topics, and many remain confined to vertical improvements within a single technical line. What is still lacking is a horizontal comparison, grounded in small-object scenarios, of image-level, feature-level, and output-level alignment strategies, together with a careful discussion of their scope and limitations. This paper presents, for the first time, a systematic review of domain-adaptive small-object detection for UAV imagery. We establish a new methodological taxonomy that organizes existing work into three categories: image-level alignment at the input layer, feature-level alignment at the representation layer, and self-training or pseudo-label strategies at the output layer. We also compare, in quantitative terms, the gains delivered by different adaptation paradigms for small-object detection. Finally, we summarize major application scenarios, identify the current limitations of the field, and outline promising directions such as source-free adaptation and lightweight edge deployment, thereby offering a theoretical basis and technical reference for practical UAV deployment across regions and weather conditions.

Key Words：

UAV; small-object detection; domain adaptation; feature alignment