Researchers proposed GFLNet, a lightweight single-stage detector for remote sensing images. The architecture addresses challenges posed by densely distributed targets and scale disparities in aerial imagery. The goal is to maintain high detection precision while controlling computational costs and inference latency. The team developed the system using collaborative feature modeling and knowledge-guided compression techniques.
The core design includes a Ghost-Shuffle dual fusion convolution module, named GSD-Fuse. This component facilitates efficient spatial-channel interaction to reduce computational overhead. To improve detection of small objects and multi-scale targets, the researchers integrated an adaptive receptive-field selection module inspired by LSKNet. This module operates in the deeper backbone and neck stages, adjusting contextual responses adaptively for specific image characteristics.
For practical deployment, the team applied a compression strategy called LB-Compress. This method combines layer-adaptive magnitude pruning with Bidirectional Cross-task Knowledge Distillation, abbreviated as BCKD. This process forms a compact network suitable for practical deployment.
Evaluation occurred on the DIOR and UCAS-AOD datasets. GFLNet improved mAP@0.5 over the baseline by 3.9% on DIOR and 3.4% on UCAS-AOD.
The final model contains only 2.35 million parameters and requires 5.6 GFLOPs for calculations. On a GPU, the inference speed reaches 261 frames per second. These metrics demonstrate a favorable accuracy-efficiency synergy.
Subject of Research: Development of a lightweight real-time object detector for remote sensing images
Article Title: GFLNet: A lightweight real-time detector with knowledge-guided compression in remote sensing images
Article References: Zhang, J., Mustaffa, M. R., Khalid, F., Kahar, Z. A., & Zhang, T. (2026). GFLNet: A lightweight real-time detector with knowledge-guided compression in remote sensing images. PLOS One, 21(10), e0361806. https://doi.org/10.1371/journal.pone.0361806
Image Credits: AI Generated
DOI: 10.1371/journal.pone.0361806
Keywords: remote sensing, object detection, GFLNet, lightweight model, knowledge distillation, DIOR dataset, UCAS-AOD, real-time inference, computer vision, model compression, lightweight, real-time
Cite Scienmag News
Elena Sutton. (November 10, 2026). Lightweight Detector GFLNet Improves mAP@0.5 Over Baseline in Remote Sensing. Scienmag. https://scienmag.com/lightweight-detector-gflnet-improves-map0-5-over-baseline-in-remote-sensing/
Elena Sutton. "Lightweight Detector GFLNet Improves mAP@0.5 Over Baseline in Remote Sensing." Scienmag, 10 November 2026, https://scienmag.com/lightweight-detector-gflnet-improves-map0-5-over-baseline-in-remote-sensing/. Accessed 22 November 2026.
Elena Sutton. "Lightweight Detector GFLNet Improves mAP@0.5 Over Baseline in Remote Sensing." Scienmag. November 10, 2026. https://scienmag.com/lightweight-detector-gflnet-improves-map0-5-over-baseline-in-remote-sensing/








