基于无人机多光谱的猕猴桃园冠层叶绿素含量检测方法

Detection Method of Chlorophyll Content in Canopy of Kiwifruit Orchard Based on UAV

摘要:

为实现对猕猴桃园区果树整体生长健康状况的快速、大规模监测，以猕猴桃园冠层叶片为研究对象，基于无人机拍摄果园多光谱图像，然后利用Pix4Dmapper软件拼接多光谱图像，获取果园的正射影像图，并进行辐射校正。切分正射影像为420个区域图像作为样本，采用最大类间方差法（Otsu）分割样本图像的冠层叶片与土壤背景，并实测每个样本的冠层SPAD值，构建冠层叶片多光谱数据集。采用箱线图法对数据集进行异常值检测，剔除异常样本；然后利用多光谱图像多通道的数据特点，提取图像的相邻通道变化率和23种常用植被指数，以及二者组合作为样本特征值，接着利用CARS、LARS、IRIV等3种特征筛选算法优选特征，分别结合偏最小二乘回归（PLSR）、支持向量回归（SVR）、岭回归（RR）、多元线性回归（MLR）和极限梯度提升树（XGBoost）、最小绝对收缩和选择算子回归（Lasso）、随机森林回归（RFR）、高斯过程回归（GPR）等8种方法构建模型，识别猕猴桃园冠层SPAD值；最后对比分析以不同样本特征构建的24个模型的性能，实验结果表明：以相邻通道变化率为特征建立的模型中，GPR模型性能最好，R2、RMSE分别为0.770、3.044；以植被指数和相邻通道变化率组合特征建立的模型中，GPR模型性能也最好，R2、RMSE分别为0.783、2.957；以植被指数为数据特征建立的XGBoost模型性能最优，R2、RMSE分别为0.787、2.933；因此基于无人机遥感的智能检测模型能够对果园冠层叶绿素含量进行准确评估。

Abstract:

Digitalization and intelligence play a crucial role in facilitating the high-quality development of the kiwifruit industry. Unlike other fruit trees, kiwi trees are vine plants that require abundant mineral nutrients during their key growth period. Inadequate management can easily lead to nutrient deficiencies, which not only affect the health of the trees but also impact the yield and quality of kiwis. Therefore, real-time monitoring of tree growth health is essential. To achieve fast and large-scale monitoring of overall growth and health in kiwi orchards, the drone was used to capture multispectral images of orchards, and then Pix4Dmapper software was utilized to splice UAV multispectral images for orthophoto maps and radiation correction on canopy leaves. The segmented orthophoto images were used as samples from 420 regions. The maximum inter-class variance (Otsu) method was employed to segment canopy leaves from soil backgrounds in the sample images, enabling measurement of canopy SPAD values for constructing a multispectral dataset. Firstly, outliers within the dataset were detected by using box plot analysis and subsequently removed as abnormal samples. Next, based on data characteristics derived from multi-channel images, feature values such as change rates between adjacent channels and 23 kinds of common vegetation indices were extracted, as well as their combination, to serve as sample feature values. Then three feature screening algorithms, including CARS, LARS, and IRIV were applied to optimize these features accordingly. Eight modeling methods, partial least square regression (PLSR), support vector regression (SVR), ridge regression (RR), multiple linear regression (MLR), extreme gradient boosting (XGBoost), least absolute shrinkage and selection operator regression (Lasso), random forest regression (RFR), and Gaussian process regression (GPR), were employed to construct models for identifying canopy chlorophyll content in macaque peach orchards. Finally, the performance of the 24 models constructed with different sample features was compared and analyzed. The experimental results showed that GPR model had the best performance among the models based on the change rate of adjacent channels, R2 and RMSE were 0.770 and 3.044, respectively. Among the models based on the combination of vegetation index and adjacent channel change rate, GPR model also had the best performance, R2 and RMSE were 0.783 and 2.957, respectively. The XGBoost model based on vegetation index was the best among all models, R2 and RMSE were 0.787 and 2.933, respectively. Consequently, the intelligent detection model utilizing UAV remote sensing enabled accurate assessment of orchard canopy chlorophyll content while facilitating analysis of orchard health status to provide decision support for subsequent intelligent orchard management.