| 引用本文: | 邢明杰,徐兴倩,徐伟恒,等.基于优选光谱特征的红壤水分反演模型研究[J].灌溉排水学报,2025,44(11):70-79. |
| XING Mingjie,XU Xingqian,XU Weiheng,et al.基于优选光谱特征的红壤水分反演模型研究[J].灌溉排水学报,2025,44(11):70-79. |
|
| 摘要: |
| 【目的】探索基于多光谱遥感技术高效监测土壤水分的方法。【方法】针对云南红壤区土壤水分遥感反演研究匮乏问题，基于无人机多光谱影像（G、R、RE、NIR波段）与同步田间采样数据，融合22种经典及改进光谱指数，结合皮尔逊相关系数（Pearson correlation coefficients, Pccs）、变量投影重要性分析（Variable importance in projection, VIP）和灰色关联度分析（Grey relational analysis, GRA）3种算法筛选敏感特征变量，分别采用随机森林（Random forest, RF）、反向传播神经网络（Back propagation neural network, BPNN）、支持向量回归（Support vector regression, SVR）及轻量梯度提升机（Light gradient boosting machine, LightGBM）模型反演红壤水分。【结果】VIP算法优选的6个关键光谱变量显著提升了模型计算效率；在4种机器学习模型中，BPNN兼具鲁棒性与泛化能力，其中VIP-BPNN模型精度最高（验证集决定系数R2=0.72，均方根误差RMSE=3.36%，相对预测性能偏差RPD=1.90），而RF模型虽然训练集R2达0.94，但测试集R2仅为0.56，存在过拟合现象。【结论】基于VIP算法与BPNN的多光谱反演模型可更好地反映红壤水分时空分布，为云南红壤区动态监测提供高精度技术支撑。未来可以融合更多波段与参数以增强模型泛化能力，进一步推动智慧农业与生态管理决策的科学化。 |
| 关键词: 多光谱；红壤水分；变量筛选；反演模型 |
| DOI：10.13522/j.cnki.ggps.2025171 |
| 分类号: |
| 基金项目: |
|
| Estimating red soil moisture using optimized spectral indices and machine learning |
|
XING Mingjie, XU Xingqian, XU Weiheng, WANG Er, ZHU Xiang, ZHAO Lin
|
|
1. College of Water Conservancy, Yunnan Agricultural University, Kunming 650201, China;
2. International College, Yunnan Agricultural University, Kunming 650201, China;
3. College of Big Data and Intelligent Engineering, Southwest Forestry University, Kunming 650224, China
|
| Abstract: |
| 【Objective】Efficient monitoring of soil moisture at large scales is required for optimizing water resource management and smart agriculture, particularly in red soils where water retention is low and irrigation efficiency is limited. This paper investigates the feasibility of using multispectral remote sensing to indirectly measure the moisture content in red soils. 【Method】The study area was in Yunnan province. Using unmanned aerial vehicle (UAV) multispectral images (green, red, red-edge, and near-infrared bands) and moisture data measured in the field, we selected 21 classical and improved spectral indices to construct an inversion model. Sensitive indices were screened using three algorithms: the Pearson correlation coefficient (PCC), variable importance in projection (VIP) and grey relational analysis (GRA). Four machine learning models: random forest (RF), back propagation neural network (BPNN), support vector regression (SVR), and light gradient boosting machine (Light-GBM) were used to estimate soil moisture content using the optimized indices.【Result】The VIP algorithm screened out six optimized spectral variables, which significantly improved computational efficiency. Among the four machine learning models we compared, the BPNN was the most robust and general. The combination of VIP and BPNN was the most accurate, and the statistical metrics of its comparison with measured field data were R2 = 0.72, RMSE = 0.03 and RPD = 1.90. The R2 of the RF model was 0.94 in the training set, but was reduced to 0.56 in the test set, indicating overfitting.【Conclusion】The multispectral inversion model using VIP and BPNN effectively captured the spatiotemporal distribution of red soil moisture in the study area. When combined with additional spectral bands and environmental parameters, this model can be applied in smart agriculture and ecological management. |
| Key words: multispectral; red soil moisture; variable screening; inversion model |
