基于多模态遥感图像融合的檀香树幼苗精准检测

Accurate Detection of Sandalwood Seedlings Based on Multi‑modal Remote Sensing Image Fusion

摘要:

檀香树幼苗在生长过程中依赖伴生植物，且常与其他作物混种，导致其生长环境复杂，难以实现精准检测，进而影响檀香树生产力的有效评估。为解决这一问题，本研究提出一种基于多光谱图像多模态融合与改进YOLO v8n的轻量级检测算法(YOLO v8n improved for sandalwood plant seedlings detection,YOLO?SPS)。利用SwinFusion模型融合檀香树幼苗的近红外与可见光遥感图像，增强其纹理与颜色特征;基于YOLO v8n模型在骨干网络中引入结合空间重构与通道重构的模块（Cross stage partial network fusion_spatial and channel reconstruction convolution，C2f_SCConv），提升特征提取能力，在颈部网络中加入部分自注意力机制（Partial self?attention，PSA），优化全局感知能力并降低计算成本，采用Shape?IoU作为边界框损失函数，进一步提高检测精度。试验结果表明，YOLO?SPS模型精确率、召回率和平均精度均值分别达到92.8%、93.2%和95.9%，较原YOLO v8n模型性能显著提升，且优于YOLO v5s、YOLO v6n、YOLO v7?tiny、YOLO v9?t、YOLO v10n等5个主流检测模型。研究结果为檀香树幼苗期精准监测提供了有效的技术支持。

Abstract:

Sandalwood seedlings rely on companion plants during their growth, which are often interplanted with other crops, resulting in a complex growth environment that makes accurate detection challenging, thereby affecting the effective assessment of sandalwood productivity. To address this issue, a lightweight detection algorithm, (YOLO v8n improved for sandalwood plant seedlings detection,YOLO?SPS), was proposed based on multispectral image fusion and an improved YOLO v8n. The SwinFusion model was utilized to fuse near?infrared and visible light remote sensing images of sandalwood seedlings, enhancing their texture and color features. The cross stage partial network fusion_spatial and channel reconstruction convolution module, which combined spatial and channel reconstruction, was introduced into the backbone network of the YOLO v8n model to improve feature extraction capabilities. The partial self?attention (PSA) mechanism was incorporated into the neck network to optimize global perception and reduce computational costs. Finally, Shape?IoU was adopted as the bounding box loss function to further enhance detection accuracy. Experimental results demonstrated that the YOLO?SPS model achieved precision, recall, and mean average precision of 92.8%, 93.2%, and 95.9%, respectively, significantly outperforming the original YOLO v8n model and surpassing five mainstream detection models, including YOLO v5s, YOLO v6n, YOLO v7?tiny, YOLO v9?t, and YOLO v10n. The research result can provide effective technical support for the precise monitoring of sandalwood seedlings during their early growth stages.