LLM-Enhanced Architecture Design for Lightweight Small-Object Detection in Low-Altitude Remote Sensing

Abstract

Low-altitude remote sensing (LARS) is increasingly used in urban monitoring, emergency response, environmental inspection, and precision agriculture. These applications require detectors that can accurately perceive dense and tiny targets from high-resolution aerial imagery while satisfying strict edge-deployment constraints. However, LARS imagery often contains small objects, large scale variation, complex backgrounds, occlusion, motion blur, and domain shifts across scenes, making lightweight object detection difficult. Meanwhile, manually designed detectors usually depend on expert experience and are difficult to adapt to diverse low-altitude remote sensing scenarios. To address these challenges, this article proposes LPNAS-Det, a large language model (LLM)-enhanced architecture planning framework for lightweight small-object detection in low-altitude remote sensing. Instead of using the LLM to directly generate detector architectures, LPNAS-Det treats the LLM as a scene-aware search planner. Specifically, measurable LARS characteristics, including small-object ratio, object density, scale variation, occlusion level, motion blur, background complexity, and deployment strictness, are first converted into detector-level requirements. The LLM then maps these requirements to adaptive search policies over backbone operators, feature-resolution hierarchies, feature-fusion structures, attention modules, detection heads, and localization losses. To improve reliability, the generated policies are projected onto the legal search space and smoothed to maintain exploration. Furthermore, a feedback-driven replanning mechanism returns Pareto-front evolution, detection performance, model complexity, invalid architecture ratio, and population diversity to the LLM planner, enabling dynamic policy adjustment during the search process. Experiments on VisDrone2019 and UAVDT show that LPNAS-Det achieves favorable accuracy–efficiency tradeoffs compared with mainstream lightweight detectors, low-altitude small-object detectors, and representative search-based methods. Additional experiments on WeedMap further demonstrate its applicability to agricultural low-altitude remote sensing. The results indicate that LLM-enhanced architecture planning provides a practical way to improve perception performance, lightweight deployment, and cross-scene adaptability for LARS applications.