| 王宏乐,刘大存,叶全洲,梁振伟,张雅朵.基于无人机遥感影像的柑橘产量估测方法[J].中国南方果树,2026,55(3): |
| 基于无人机遥感影像的柑橘产量估测方法 |
| Estimation of Citrus Yield Based on UAV Remote Sensing Images |
| 投稿时间：2025-05-16 修订日期：2025-06-11 |
| DOI：10.13938/j.issn.1007-1431.20250228 |
| 中文关键词: 柑橘 冠层 遥感影像 产量 深度学习 无人机 |
| 英文关键词:citrus canopy remote sensing images yield UAV |
| 基金项目:深圳市科技计划资助(CJGJZD20210408092401004)和(KCXFZ20240903093800002)、广东省现代农业产业园项目(GDSCYY2022-046)※ |
|
| 摘要点击次数: 132 |
| 全文下载次数: 0 |
| 中文摘要: |
| 【目的】本研究旨在基于利用深度学习技术,从无人机多光谱影像中提取融合树冠光谱特征和株高特征,探索柑橘产量预估方法,为快速准确获取柑橘产量提供理论和实践基础。【方法】在广东韶关南雄沃柑基地,随机选取99株柑橘树样本,人工调查统计每一柑橘单株挂果数量。通过多光谱无人机获得多光谱遥感影像数据及数字表面模型(Digital Surface model, DSM)。通过Python处理 DSM获得株高影像。将无人机获取的不同波段 RGB、多光谱影像与株高影像进行分组,分别采用 ResNet18、ResNet34 和 MobileNetV2 进行训练,随后对模型预测精度进行比较分析, 筛选较优产量估测模型。【结果】 基于RGB影像的ResNet18、ResNet34和MobileNetV2模型的估产精度分别为89.93%、89.72%和83.97%,相关系数(R2)分别为0.7121、0.6874和0.4137。当训练数据由 RGB 波段和株高影像组成时,ResNet18、ResNet34和MobileNetV2的估产精度分别为90.29%、90.36%和83.63%,R2分别提高至0.7500、0.7226和0.5522。当训练数据由 RGB 波段和多光谱波段组成时,ResNet18、ResNet34和MobileNetV2的估产精度分别为91.14%、91.14%和83.79%,R2分别为0.7714、0.7857和0.8297。当训练数据由 RGB、多光谱和株高数据组成时,ResNet18、ResNet34和MobileNetV2的估产精度分别为91.43%、91.06%和86.34%；R2值分别为0.7971、0.8297和0.5667。【结论】以RGB影像组合株高影像及多光谱影像数据建立的单株挂果量预测模型具有较高的预测精度及相关系数；无人机多光谱影像和株高影像数据融合,使用深度学习方法进行柑橘产量预估,最佳模型平均精度为91.06%,R2为0.8297,可以用于快速准确预估柑橘单株产量。本研究为利用无人机多光谱影像进行柑橘产量快速精准预估的方法提供了理论与实践依据。 |
| 英文摘要: |
| 【Objective】 This study aims to extract and fuse canopy spectral features and plant height features from UAV multispectral image by using deep learning techniques, and provide a theoretical and practical basis for the rapid and accurate acquisition of citrus yields. 【Methods】The study was carried out in a Citrus (cultivar: Orah) garden located in Nanxiong, Shaoguan, Guangdong province. Number of 99 citrus plants were randomly selected and the fruit number of each plant was investigated. Multispectral remote sensing images and the Digital Surface Model (DSM) were obtained through a multispectral unmanned aerial vehicle (UAV). The plant height image was obtained by processing the DSM using Python script. The UAV-derived bands including RGB, multispectral and the plant height image were grouped and trained using ResNet18, ResNet34, and MobileNetV2, respectively. The accuracies of models were compared and analyzed, to screen an optimal model for citrus yield estimation. 【Results】 When the training images were RGB images, the yield estimation accuracies of ResNet18, ResNet34, and MobileNetV2 were 89.93%, 89.72%, and 83.97%, respectively; and the corresponding correlation coefficient (R2) values were 0.7121, 0.6874, and 0.4137. When the training images bands composed of RGB and plant height bands, the accuracies of ResNet18, ResNet34, and MobileNetV2 were 90.29%, 90.36%, and 83.63%, respectively; and their R2 values were 0.7500, 0.7226, and 0.5522. When the training images bands composed of RGB and multispectral bands, the accuracies were 91.14%, 91.14%, and 83.79% for ResNet18, ResNet34, and MobileNetV2, respectively; and their R2 values were 0.7714, 0.7857, and 0.8297. When the training images bands composed of RGB, multispectral and plant height bands, the accuracies of ResNet18, ResNet34, and MobileNetV2 were 91.43%, 91.06%, and 86.34%, respectively; and their R2 values were 0.7971, 0.8297, and 0.5667. 【Conclusion】 Models developed by adding plant height and multispectral bands into RGB image showed an improved yield prediction in accuracy and correlation coefficients. The accuracy of the optimal model was 91.06%, and the R2 was 0.8297 for citrus yield estimation, by using deep learning method training images composed of UAV multispectral and plant height bands. This study provides a theoretical and practical basis for the rapid and accurate method for citrus yield estimation using UAV multispectral images. |
| 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
