基于无人机遥感特征选择的冬小麦氮营养指数估测模型研究

Estimation of Nitrogen Nutrition Index in Winter Wheat Based on UAV Remote Sensing and Feature Selection

摘要:

氮营养指数(Nitrogen nutrition index, NNI)是表征冬小麦氮营养状态及辅助精准施肥决策的关键指标。尽管无人机遥感技术为大面积快速获取NNI 提供了有效手段,但如何科学融合多维遥感特征、优化特征筛选流程并构建高精度估测模型,仍是当前提升NNI 监测精度的难点。本研究基于无人机遥感数据,提取并组合植被指数特征和冠层结构特征,提出了一种基于Shapley additive explanations(SHAP)算法与递归消除的特征选择方法,该方法量化了每个特征对模型预测的贡献,能更精确地筛选出关键特征。使用PLSR(Partial least squares regression)、KNN(K-nearest neighbors)、GPR(Gaussian process regression)、XGBoost(Extreme gradient boosting)、AutoGluon共5种算法构建冬小麦氮营养指数估测模型,并进行有效评估。结果表明:特征融合与筛选显著提升了估测精度,从融合特征中筛选出的23个核心特征在模型表现上最优,精度由大到小为融合特征、植被指数特征、冠层结构特征的规律,证明了多维特征互补及筛选优化在提升NNI 监测性能中的关键作用;在模型精度对比中,AutoGluon在3种特征组合下均表现最佳,最高R2达到0.78,显示出良好的稳定性与泛化能力。本研究识别了氮营养反演的有效特征,提升了拔节期冬小麦NNI的估测精度,为精准施肥提供了科学依据。

Abstract:

The nitrogen nutrition index (NNI) served as a critical indicator for characterizing the nitrogen status of winter wheat and supporting precision fertilization decision-making. Although UAV-based remote sensing technology provided an effective means for the rapid, large-scale acquisition of NNI, scientifically integrating multidimensional remote sensing features, optimizing feature selection workflows, and constructing high-precision estimation models remained significant challenges in improving NNI monitoring accuracy. Multi-source remote sensing data was utilized to extract and combine vegetation index features and canopy structural features, and a feature selection method was proposed based on Shapley additive explanations (SHAP) and recursive feature elimination (RFE). By integrating SHAP values with the recursive elimination strategy, it was able to more accurately identify key features. Five algorithms, i. e., partial least squares regression (PLSR), K-nearest neighbors (KNN), Gaussian process regression (GPR), extreme gradient boosting (XGBoost) and AutoGluon were employed to construct nitrogen nutrition index estimation models for winter wheat, with comprehensive model evaluation conducted. The results showed that the integration and selection of features significantly enhanced the estimation accuracy. Among the configurations tested, the 23 core features selected from the fused feature set achieved the optimal model performance. The accuracy followed the hierarchical order of fused features, vegetation indices, and canopy structural features, which demonstrated that the complementarity of multidimensional features and the optimization of feature selection played a crucial role in improving NNI monitoring performance. In the comparison of model accuracies, AutoGluon performed the best across all three feature combinations, reaching a maximum R2 of 0.78, which exhibited robust stability and generalization capability. The research result identified effective features for nitrogen inversion and improved the estimation accuracy of NNI during the jointing stage of winter wheat, providing a scientific basis for precision fertilization and field management.