基于改进U-Net的无人机可见光影像植被精细分割研究

Research on Vegetation Segmentation in UAV Imagery Based on Improved U-Net

摘要: 利用高分辨率无人机可见光影像实现植被类型的自动、精细分割，对于生态环境监测与量化研究具有重要意义。然而，可见光影像光谱信息有限，加之植被类型间光谱特征相似(特别是乔木、灌木与草地)，导致精细分割精度面临挑战。本文提出一种改进的U-Net语义分割模型，旨在提升植被精细分割的精度与模型泛化能力。主要改进包括：1) 采用在ImageNet上预训练的ResNet34网络替换原始编码器，以增强多层次特征提取能力；2) 在跳跃连接中引入空间与通道挤压激励注意力模块，实现特征的自适应融合；3) 设计一种结合交叉熵损失、Dice损失与Focal损失的复合损失函数，以优化训练过程并聚焦难分样本。本研究在自主构建的无人机影像数据集上进行实验验证。实验采用五折交叉验证与OneCycleLR学习率调度策略。结果表明，本文模型在B测试集上的总体精度达到90.2%，较原始U-Net提升6.3个百分点，其中对灌木和草地的分割精度提升尤为显著。本研究为基于低成本可见光数据的植被精细识别提供了一种有效的深度学习解决方案。

Abstract: Realizing the automatic and fine-grained segmentation of vegetation types using high-resolution UAV visible imagery is of great significance for ecological environment monitoring and quantitative research. However, the limited spectral information of visible imagery, coupled with the similar spectral characteristics among different vegetation types (especially arbor, shrub, and grassland), poses challenges to the accuracy of segmentation. This study proposes an improved U-Net semantic segmentation model aiming to enhance the segmentation accuracy and model generalization ability. The main improvements include: 1) replacing the original encoder with the ResNet34 network pre-trained on ImageNet to strengthen multi-level feature extraction capability; 2) introducing spatial and channel squeeze-and-excitation (SE) attention modules into skip connections to achieve adaptive feature fusion; 3) designing a hybrid loss function combining cross-entropy loss, Dice loss, and Focal loss to optimize the training process and focus on hard-to-segment samples. Experimental verification was conducted on a self-constructed UAV imagery dataset, adopting five-fold cross-validation and the OneCycleLR learning rate scheduling strategy. The results show that the overall accuracy (OA) of the proposed model on Test Set B reaches 90.2%, improving by 6.3 percentage points compared with the original U-Net. Particularly, the segmentation accuracy of shrubs and grassland is significantly enhanced. This study provides an effective deep learning solution for fine-grained vegetation identification based on low-cost visible light data.