| 符欣彤,常庆瑞,张佑铭,张子娟,郑智康,李铠.基于Stacking集成学习的猕猴桃叶片叶绿素含量估算[J].干旱地区农业研究,2023,(4):247~256 |
| 基于Stacking集成学习的猕猴桃叶片叶绿素含量估算 |
| Estimation of kiwifruit leaf chlorophyll content based on Stacking ensemble learning |
| |
| DOI：10.7606/j.issn.1000-7601.2023.04.26 |
| 中文关键词: 猕猴桃 叶绿素含量 高光谱 Stacking集成学习 |
| 英文关键词:kiwifruit chlorophyll content hyperspectral Stacking ensemble learning |
| 基金项目:国家863计划项目(2013AA102401-2) |
|
| 摘要点击次数: 681 |
| 全文下载次数: 531 |
| 中文摘要: |
| 叶绿素含量能有效表征植物光合作用强度，是反映植物生长状况的重要参量之一。以秦岭北麓壮果期猕猴桃叶片为研究对象，分别测定其叶绿素含量和光谱反射率，通过分析380～1 000 nm范围内高光谱参数与叶绿素含量的相关性，筛选出估测模型的输入特征，选择随机森林、极限梯度提升树、K-近邻、LightGBM算法和岭回归作为基模型，线性回归作为元模型，建立基于Stacking集成学习的猕猴桃叶片叶绿素含量估算模型，并通过网格搜索和交叉验证提高模型泛化能力，将Stacking模型与多个单一模型进行比较。结果表明：（1）不同叶绿素含量的猕猴桃叶片高光谱反射率变化趋势基本一致，在380～1 000 nm范围内呈现“一峰两谷一平台”的特点；（2）各高光谱参数与猕猴桃叶片叶绿素含量相关性较好，优化光谱指数和传统光谱指数中与叶绿素含量相关性最高的分别是比值光谱指数（RSI′581,438,r=0.947）和红边位置（r=0.914）；（3）与多个单一模型相比，Stacking集成模型的估算精度最高（R2=0.807, MAE=0.334, RMSE=0.136），同时，其相对预测偏差（RPD=7.443）明显高于其他模型，具有极好的预测能力。本研究为快速、无损、精确地获取猕猴桃叶片叶绿素含量提供了新思路。 |
| 英文摘要: |
| Chlorophyll content can effectively represent the intensity of plant photosynthesis which is one of the important parameters reflecting plant growth status. The chlorophyll content and spectral reflectance of kiwifruit leaves at the strong fruit stage at the northern foot of the Qinling Mountains were measured. By analyzing the correlation between hyperspectral parameters and chlorophyll content in the range of 380~1 000 nm, the input characteristics of the estimation model were screened out. Random forest, extreme gradient boosting, K-nearest neighbor, light gradient boosting machine, and ridge regression were selected as the base models and linear regression as the meta model to establish the estimation model of chlorophyll content in kiwifruit leaves based on Stacking ensemble learning. The generalization ability of the model was improved through grid search and 5-fold cross, and the Stacking model was compared with multiple single models. The results showed that: (1) The variation trend of hyperspectral reflectance of kiwifruit leaves with different chlorophyll contents was basically the same, showing the characteristics of “one peak, two valleys and one platform” in the range of 380~1 000 nm. (2) The correlation between hyperspectral parameters and chlorophyll content in kiwifruit leaves was good. In the optimized spectral index and the traditional spectral index, the highest correlation with chlorophyll content was ratio spectral index（RSI′581,438,r=0.947) and red edge position (r=0.914); (3) Compared with multiple single models, the Stacking ensemble model had the highest estimation accuracy (R2=0.807, MAE=0.334, RMSE=0.136). At the same time, its relative prediction deviation (RPD=7.443) was significantly higher than other models, which had excellent prediction ability. This study provides a new idea for obtaining the chlorophyll content of kiwifruit leaves quickly and accurately. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| | |