MGSC-Net: Multigranularity Semantic Aggregation and Scale-Calibration for Small Object Detection in UAV Remote Sensing Imagery

Uncrewed aerial vehicle remote sensing has become a pivotal technology for high-resolution Earth observation tasks, such as urban traffic monitoring and emergency rescue. However, processing these images faces severe challenges due to the complex acquisition conditions inherent to aerial platforms. Primarily, the large span of flight altitudes and variable gimbal pitch angles result in significant fluctuations in ground sampling distance (GSD), leading to a “scale discontinuity” where the projected sizes of same-class objects vary drastically. This intraclass geometric deformation causes severe semantic misalignment across feature pyramid layers. Secondarily, complex land cover textures (e.g., vegetation and shadows) in high-altitude top–down views easily submerge the spectral and spatial features of tiny targets. To address these issues specific to aerial sensing, this article proposes a high-performance detection framework named MGSC-Net. First, to mitigate resolution degradation, a lossless attentional downsampling (LAD) module is designed in the backbone, leveraging space-to-depth conversion and coordinate attention to preserve high-frequency geospatial details. Second, a scale feature calibration (SFC) module is introduced to align cross-scale representations. By employing a dynamic gating mechanism, SFC adaptively calibrates multilayer features based on semantic content, effectively eliminating misalignment caused by GSD variations. Furthermore, a multigranularity semantic aggregation (MGSA) module is integrated into the detection head. By constructing parallel streams for spatial details and local/global contexts, MGSA effectively decouples target features from complex environmental clutter via an attention-guided fusion mechanism. Extensive experiments on the VisDrone2019 and USOD benchmarks demonstrate the robustness of MGSC-Net in complex remote sensing scenarios, achieving mAP50 scores of 39.3.