基于敏感变量筛选的多光谱植被含水率反演模型研究

Multispectral Vegetation Water Content Inversion ModelBased on Sensitive Variable Filtering

植被含水率是农田生态系统敏感性的重要表征,为提高近地遥感植被含水率反演效率和精度,基于无人机多光谱影像数据,提取苜蓿、玉米2种植被覆盖的光谱反射率,在此基础上引入红边波段计算改进光谱指数。将5种光谱反射率及25个光谱指数利用变量投影重要性(Variable importance in projection, VIP)分析、灰色关联度(Gray relational analysis, GRA)分析与皮尔逊(Person)相关性分析进行筛选,并建立基于反向神经网络(Back-propagation neural network, BPNN)、偏最小二乘法(Partial least squares regression, PLSR)、支持向量回归(Support vector regression, SVR)和随机森林(Random forest, RF)4种机器学习模型,以确定不同作物覆盖下的最佳植被含水率反演模型。结果表明,3种筛选算法中VIP和GRA的模型精度明显优于Person相关性分析,且反演结果波动较小;在4种机器学习算法中,SVR算法在非线性问题中相较于BPNN、PLSR、RF算法具有较强的解析能力和模型鲁棒性,验证集决定系数R~2达到0.77以上,其结果能较真实反映植被含水率;两种样地基于GRA的植被含水率反演模型精度最高,苜蓿覆盖地GRA-SVR验证集R~2达0.889,RMSE为0.798%,MAE为0.533%;玉米覆盖地反演结果验证集R~2为0.848,RMSE为0.668%,MAE为0.542%。研究结果可为植被含水率的快速、精准反演提供理论依据。

Vegetation moisture content is an important characterization of the sensitivity of farmland ecosystem.The spectral reflectance of two vegetation covers,alfalfa and corn were extracted,based on the UAV multispectral image data,and on the basis of which the red-edge band was introduced to calculate the improved spectral indices in order to increase the efficiency and accuracy of the inversion of vegetation water content by near-earth remote sensing.A back-propagation neural network(BPNN)was created after the five spectral bands and 25 indices were filtered by using the variable importance in projection(VIP),gray relational analysis(GRA),and Pearson's correlation analysis.To find the optimum inversion model for vegetation water content under various crop covers,back-propagation neural network,partial least squares regression(PLSR),support vector regression(SVR),and random forest(RF)were used.The findings indicated that,among the three screening algorithms,the accuracy of the models following GRA and VIP was significantly higher than that of Pearson's correlation analysis,and the inversion results were less volatile.Among the four machine learning algorithms,the SVR algorithm had a stronger nonlinear problem resolution ability and model robustness than BPNN,PLSR,and RF algorithms.In the nonlinear problem,the SVR algorithm outperformed the BPNN,PLSR,and RF algorithms in terms of analytical ability and model robustness.The validation set coefficient of determination R2 reached above 0.77 and its results can offer more accurate feedback on vegetation water content.The GRA-SVR based inversion model for vegetation water content had the highest accuracy in the two sample sites.The GRA-SVR validation set R2 of alfalfa cover reached 0.889,RMSE of 0.798%,and MAE of 0.533%;the inversion result validation set R2 of corn cover was 0.848,RMSE of 0.668%,and MAE of 0.542%.The research results can provide a theoretical basis for rapid and accurate inversion of vegetation water content.