基于支持向量机回归算法的小麦叶面积指数高光谱遥感反演

Wheat leaf area index inversion with hyperspectral remote sensing based on support vector regression algorithm

为给小麦田间管理提供基础数据,利用高光谱指数实现了小麦冠层叶面积指数(LAI)值的估测。在21种高光谱指数中筛选出了与LAI值相关性最强的指数OSAVI,建立了小麦LAI值反演的最小二乘支持向量回归(LS-SVR)模型。分析表明,模型校正集决定系数(C-R2)与预测集决定系数(P-R2)分别达0.851与0.848,可实现小麦LAI值的精确反演,且对LAI值较高与较低的样本均具备良好的预测能力,可有效避免冠层郁闭度等因素对估测结果的影响。利用LS-SVR模型与OMIS影像实现了小麦LAI遥感专题图的制作,其填图结果与地面实测值拟合模型R2达0.774,RMSE仅为0.476,2组数据具有较高的相似度。结果表明:可利用高光谱指数实现小麦冠层LAI值信息的准确获取,且OSAVI系反演建模的优选指数,LS-SVR为建模的优选算法。该研究可为小麦等农作物的长势评估提供参考。

Determination of crops＇ leaf area index （LAI） is of great significance for growth monitoring, water-fertilizer regulation and yield assessment. For the sake of providing basic data for wheat field management, estimation of LAI value of wheat canopy was conducted by using hyperspectral indices. The optimization of Soil-adjusted Vegetation Index （OSAVI） which is the strongest correlation with LAI was selected from 16 kinds of existing hyperspectral indices like GREEN-NDVI and 5 kinds of newest established hyperspectral indices like FD730 , and linear model for wheat LAI inversion was established by adopting the Least Squares Method algorithm. The analysis results showed that the calibration set decision coefficient （C-R^2 ） and prediction set decision coefficient （P-R^2 ） of the model reached 0.832 and 0.825 respectively, the Root Mean Square Error of Calibration set （RMSEC） and the Root Mean Square Error of Prediction set （RMSEP） were 0.478 and 0.461 correspondingly, so the accurate inversion of wheat LAI could been realized. To further improve inversion precision, the model was optimized by using the Least Squares Support Vector Regression （LS-SVR）. In comparison with linear model, the coefficients of C-R^2 and P-R^2 reached 0.851 and 0.848 respectively, obviously, higher than the ones of linear model. In the meantime, RMSEC and RMSEP were 0.467 and 0.441 correspondingly, lower than the ones of linear model. The facts also demonstrated that the LS-SVR model was better than linear model for inversion. In order to analyze prediction ability of OSVAI with regard to different LAI samples, comparative analysis was implemented between OSVAL index and the indices such as GREEN-NDVI. The results indicated that OSVAI model built had good prediction ability for the higher LAI value samples and the lower LAI value samples, and meanwhile it could also avoid influencing the result of estimation by the canopy density effectively. Finally, remote sensing thematic map of wheat LAI was achieved by using the LS-SVR model with the OMIS images. By comparing the map result with the ground measurement, the R^2 value of fitting model was 0.774, the RMSE was only 0.476, which proved that higher similarity existed in the two sets of data. The results indicated that wheat canopy LAI information could be acquired accurately by using hyperspectral indices, and OSAVI was optimal index for inversion modeling, LS-SVR was the optimization algorithm for modeling. The study can provide a reference for crops growth assessment such as wheat.