From UAV imagery to mapping: Detecting sunflower heads in fields using a novel lightweight deep learning network

Abstract

Efficient and non-invasive field-scale detection and localization of sunflower heads (SHs), together with spatial distribution mapping, can support pre-harvest yield prediction, optimization of mechanical harvesting, field management, and high-throughput phenotyping. Unmanned aerial vehicle (UAV) imagery, with its low cost and high spatiotemporal resolution, makes such field-scale monitoring practically feasible. However, accurately detecting and mapping individual SHs from high-resolution UAV images remains challenging, especially under resource-constrained computing environments. To address this, we used UAV RGB imagery to construct a sunflower head detection dataset covering both flowering and maturity stages. Furthermore, a lightweight deep learning network, built upon YOLOv8n improvements, was proposed to enable efficient head detection and mapping. First, the DWCSP module is introduced, utilizing depthwise convolution and multi-branch feature fusion for feature extraction, thereby significantly reducing the network complexity. Additionally, a lightweight detection head integrating partial convolution was designed to further accelerate inference speed, and the WIoU loss function was adopted to enhance detection performance. Experimental results revealed that, when compared to the baseline, the computational complexity and parameters of the proposed model were reduced by 60.5 % and 49.5 %, respectively, with values reaching 3.2 GFLOPs and 1.52 M and a model size of only 3.1 MB. This model achieved an impressive 96.2 % [email protected]. When deployed on a CPU and the Jetson Orin Nano platform, inference speeds of 16 FPS and 67 FPS were attained, representing improvements of 33.3 % and 24.1 % over the baseline. Additionally, the model was employed to perform overlapping slice detection on UAV orthomosaic images from two sample fields, mapping individual SHs locations to geographic coordinates and generating spatial and density distribution maps of the heads. This produces an end-to-end workflow from UAV imagery to geospatial data, which provides an effective approach for pre-harvest yield estimation and analysis of agronomic variability in sunflowers, with the potential to reduce resource waste and labor demands, while providing cost-effective tools for breeding evaluation and decision-making.