| 引用本文: | 曾禧童,张 伟,付 静,等.耦合无人机高光谱与机器学习的辣椒田土壤水分动态监测研究[J].灌溉排水学报,2026,45(8):9-16. |
| Zeng Xitong,Zhang Wei,Fu Jing,et al.耦合无人机高光谱与机器学习的辣椒田土壤水分动态监测研究[J].灌溉排水学报,2026,45(8):9-16. |
|
| 摘要: |
| 【目的】基于无人机高光谱遥感和机器学习算法构建辣椒田土壤含水率动态反演模型。【方法】以贵州省典型山地的辣椒田为研究对象，基于田间小区试验，采集9期高光谱影像数据及地面土壤含水率样本，在筛选土壤含水率敏感光谱特征的基础上，结合光谱特征空间优化，采用机器学习构建辣椒田土壤含水率动态反演模型。【结果】以筛选得到的增强植被指数（EVI）、差值植被指数（DVI）、比值植被指数2（RVI2）及波段5（B5，400.878 nm）、波段128（B128，926.664 nm）等36个敏感波段作为反演模型的输入特征时，辣椒田土壤含水率反演效果最优；对比分析偏最小二乘回归（PLSR）、L2正则化回归（L2）、决策树（DT）、随机森林（RF）与类别提升（CatBoost）算法的反演表现，发现CatBoost模型的反演精度最高，其测试集决定系数（R2）为0.758、均方根误差（RMSE）为2.897、平均绝对百分比误差（MAPE）为5.902%。【结论】本研究筛选出了对辣椒田土壤水分敏感的有效光谱特征组合，通过模型比选与精度评估验证了CatBoost模型的非线性拟合能力及对山地零散灌片辣椒田土壤水分动态反演的适用性。 |
| 关键词: 辣椒；无人机遥感；高光谱；土壤含水率；CatBoost |
| DOI：10.13522/j.cnki.ggps.2026030 |
| 分类号: |
| 基金项目: |
|
| UAV hyperspectral remote sensing combined with machine learning for dynamic monitoring of soil moisture in mountainous chili pepper fields |
|
Zeng Xitong, Zhang Wei, Fu Jing, Wang Jie, Li Changjiang, He Yupu, Wan Jiawei
|
|
1. Nanjing Hydraulic Research Institute, Nanjing 210029, China;
2. Jiangsu Province Rural Water Conservancy Science and Technology Development Center, Nanjing 210029, China;
3. Guizhou Provincial Institute of Water Resources Science, Guiyang 550002, China
|
| Abstract: |
| 【Objective】Chili pepper is an important economic crop in Guizhou Province, and its healthy growth depends on a favourable soil moisture environment. Traditional soil moisture monitoring systems are costly and point-based with poor spatial representativeness. Although UAV hyperspectral remote sensing and machine learning enable soil moisture inversion, the optimal spectral feature combination and inversion algorithm for chili pepper grown in complex hilly regions remain unclear. This study aims to fill this technological gap.【Method】The field experiment was conducted in a fragmented mountainous chili irrigation area in Guizhou Province. Nine phases of UAV hyperspectral images and the corresponding soil moisture content (SMC) data were measured in situ in the field. After screening SMC-sensitive spectral features and optimizing the spectral feature space, five machine learning algorithms, including Partial Least Squares Regression (PLSR), L2 Regularization Regression (L2), Decision Tree (DT), Random Forest (RF), and Categorical Boosting (CatBoost), were used to construct the SMC inversion model. The accuracy of each model was tested against ground-truth data.【Result】The optimal model inputs were 36 SMC-sensitive spectral bands (represented by B5 at 400.878 nm and B128 at 926.664 nm) and three spectral indices: Enhanced Vegetation Index (EVI), Difference Vegetation Index (DVI), and Ratio Vegetation Index 2 (RVI2). Among all tested algorithms, CatBoost yielded the most accurate inversion on the test set, with a coefficient of determination of 0.758, a root mean square error of 2.897, and a mean absolute percentage error of 5.902%.【Conclusion】For monitoring soil moisture content in fragmented mountainous chili pepper fields, the CatBoost algorithm achieved the highest accuracy. Integrating the UAV hyperspectral technique and CatBoost model can reliably monitor soil moisture dynamics, facilitate precise water regulation and improve water use efficiency. |
| Key words: chilli pepper; UAV remote sensing; hyperspectral data; soil moisture content; CatBoost |
