Research Article | Open Access

Volume 2023 |Article ID 0123 | https://doi.org/10.34133/plantphenomics.0123

Lightweight Deep Learning Models for High-Precision Rice Seedling Segmentation from UAV-Based Multispectral Images

Panli Zhang,¹ Xiaobo Sun ,¹ Donghui Zhang,² Yuechao Yang,³ and Zhenhua Wang⁴

¹College of Engineering, Northeast Agricultural University, Harbin 150030, China.
²Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
³National Key Laboratory of Remote Sensing Information and Imagery Analyzing Technology, Beijing Research Institute of Uranium Geology, Beijing 100029, China
⁴College of Agriculture, Northeast Agricultural University, Harbin 150030, China

Abstract

Accurate segmentation and detection of rice seedlings is essential for precision agriculture and high-yield cultivation. However, current methods suffer from high computational complexity and poor robustness to different rice varieties and densities. This article proposes 2 lightweight neural network architectures, LW-Segnet and LW-Unet, for high-precision rice seedling segmentation. The networks adopt an encoder–decoder structure with hybrid lightweight convolutions and spatial pyramid dilated convolutions, achieving accurate segmentation while reducing model parameters. Multispectral imagery acquired by unmanned aerial vehicle (UAV) was used to train and test the models covering 3 rice varieties and different planting densities. Experimental results demonstrate that the proposed LW-Segnet and LW-Unet models achieve higher F1-scores and intersection over union values for seedling detection and row segmentation across varieties, indicating improved segmentation accuracy. Furthermore, the models exhibit stable performance when handling different varieties and densities, showing strong robustness. In terms of efficiency, the networks have lower graphics processing unit memory usage, complexity, and parameters but faster inference speeds, reflecting higher computational efficiency. In particular, the fast speed of LW-Unet indicates potential for real-time applications. The study presents lightweight yet effective neural network architectures for agricultural tasks. By handling multiple rice varieties and densities with high accuracy, efficiency, and robustness, the models show promise for use in edge devices and UAVs to assist precision farming and crop management. The findings provide valuable insights into designing lightweight deep learning models to tackle complex agricultural problems.