Optimizing long-range UAV detection on YOLOv8: Breaking-point distance analysis and combining adaptive tiling with AdamW optimizer

Abstract

The rapid proliferation of unmanned aerial vehicles (UAVs) has imposed stringent requirements on surveillance and early warning systems. In long-range detection scenarios, the apparent size of UAVs in images decreases significantly, leading to severe spatial information loss and degraded performance of convolutional neural network (CNN)-based detection models. This paper proposes a continuous quantitative analysis framework to model the relationship between observation distance and UAV detection performance by progressively reducing the input image resolution. Based on experimental regression analysis, a system-level breaking point is identified, representing a distance threshold at which detection performance begins to degrade sharply and exhibits nonlinear behavior. Furthermore, a solution integrating adaptive image tiling with the AdamW optimizer is proposed to ensure training stability and enhance performance in long-range scenarios. Experimental results on the YOLOv8s model show that the proposed approach improves mAP@0.5 in long-range detection by up to +24.9% while eliminating numerical instability during training on tiled data. Regression analysis identifies the system-level breaking point at D_c ≈ 2.5, providing a quantitative basis for activating adaptive image processing in real-world deployments on resource-constrained platforms.