基于支持向量机的水稻叶面积指数高光谱估算模型研究

为了研究支持向量机(SVM)对于作物农学参数高光谱估算的能力，通过大田小区试验，测定了2个品种、3个供氮水平处理的水稻在不同生长期的冠层高光谱反射率(350～2 500 nm)。依据Ladsat-5的TM传感器波段宽度，将高光谱反射率转换为10种不同的植被指数。利用所有样本的植被指数和水稻叶面积指数(LAI)，通过不同统计模型的模拟分析，依据模型的R2选取了三种相关性较高的统计关系(包括NDVI_green的指数关系、TCARI/OSAVI的乘幂关系和RVI₂的乘幂关系)。对这三种关系，通过具有不同核函数的SVM模型和相应统计模型对LAI进行估算。结果表明：所有的SVM模型都具有较低的均方根误差值，估算精度都高于相应的统计模型；基于TCARI/OSAVI的POLY核SVM具有最高的估算精度，其RMSE比相应的统计模型降低近11个百分点。因此，SVM方法用于水稻LAI高光谱估算具有良好的学习能力和鲁棒性。

The Estimation Model of Rice Leaf Area Index Using Hyperspectral Data Based on Support Vector Machine

In order to compare the prediction powers between the best statistical model and SVM technique using each VI for rice LAI, the VIs are as independent variables in statistical models and are as net inputs in SVM, and the rice LAI are as dependent variables in statistical models and are as net outputs in SVM.Hyperspectral reflectance (350 to 2 500 nm) data were recorded in two experiments involving four replicates of two rice cultivars (“Xiushui 110” and “Xieyou 9308”), three nitrogen levels (0, 120, 240 kg·ha-1 N), and with a plant density of 45 plants·m-2.The first experiment was seeded on 30 May 2004 and the second experiment on 15 June 2004.Both sets of seedlings were transplanted to the field one month later.Hyperspectral reflectance was ground-based and measured using Analytical Spectral DevicesTM and 1 meter above the rice canopy.The solar angle compared to nadir was for all measurements less than 45° and no disturbing clouds were observed.Hyperspectral reflectance was transformed to ten different vegetation indices including RVI, NDVI, NDVI_green, SAVI, OSAVI, MSAVI, MCACI, TCARI/OSAVI, RDVI and RVI₂, according to the width of TM bands of Ladsat-5.Different statistical models including linearity model, exponent model, power model and logarithm model, were analyzed using all samples’ LAI and vegetation indices.Three good relationships including exponent relationship of NDVI_green, power relationship of TCARI/OSAVI and power relationship of RV₁₂ were selected based on the R2 of models.These three relationships were used to predict the LAI of rice through SVM models with different kernel functions including an analysis of variance kernel (ANOVA), a polynomial kernel (POLY) and a radial basic function kernel (RBF), and corresponding statistical models.The results show that all SVM models have lower RMSE values and higher estimation precision than corresponding statistical models；SVM with POLY kernel function using TCARI/OSAVI has the highest estimation precision for rice LAI compared to other models, and it’s RMSE value is lower than corresponding statistical model by 11 percent points.Therefore, SVM has a high accuracy for learning and a good robustness for estimation of LAI of rice using hyperspectral data.Consequently, SVM provides a useful explorative tool for improvement of the relationships between VIs and rice LAI.