基于支撑向量回归的二端元混合像元分解

Two-endmember mixed pixel unmixing based on SVR model

针对遥感影像混合像元光谱复杂,其非线性特征,传统LSMM分解模型难以进行有效的混合像元分解的不足。通过基于SVR的二端元混合像元分解的研究,从真实遥感影像上获取典型的植被、非植被光谱信息,构造二端元混合光谱库,进行SVR模型的混合像元分解。当样本量为6%时,交叉验证获得最佳模型参数(C=1024.0和g=4.0),进一步对全部混合像元进行混合像元分解。实验结果表明:SVR分解结果 RMSE为5.95,R2为0.958,优于LSMM方法(RMSE=7.71,R2=0.932),且在各个不同真值丰度下具有更好的稳定性,证明该方法对于非线性混合光谱具有很好的学习和推广能力。此外,该方法的精度不随训练样本量的增加呈明显变化,体现出SVR在有限样本情况下能够保证高效率的训练能力。

Due to the spectral heterogeneity of mixed pixel spectral with non-linear characteristics from remote sensing, the traditional linear spectral mixture model, called as linear spectral mixed mode called as LSMM, can- not solve the mixed pixel for mapping land cover fraction effectively. In order to improve the performance of un- mixing mixed pixels, the paper introduced the support vector regression named as SVR model to address the non-linear spectral problem. SVR model owns the advantages of searching optimum balance ability with small amount training sample among complex model and study ability to achieve high application performance. In pa- per, the SVR model was carried out with five steps in a simulated experiment. First, from real remote sensing im- age, one thousand typical vegetation and bare land spectral pixels were extracted from the classified remote sens- ing image which resolution is 2.5 m, respectively. Second, the training sample set to form the different vegetation fraction pixels, such as 0%, 1%, 2%, ＂＂, 100%, was constructed from the vegetation and bare land. The simulated remote sensing images were produced for the further study. Finally, SVR model was carried out for unmixing the mixed pixels to extract the vegetation fraction. The two sensitive parameters, C and g, to ensure the SVR perfor- mance which was determined by grid-regression method. In order to validate the influence of unmixing accuracy with different amount training sample, 1%, 2%,…, 10% sample SVR experiment were also carried out. The high- est regression accuracy and most optimum model parameters produced by SVR regression model was 6% sample amount. Hence, small sample amount is enough for mixed pixels unmixing using SVR method, which is better than that of LSMM which need amount of endmembers deriving from remote sensing image to obtain the average spectrum of each land cover to estimate land cover fraction. SVR method can obtain higher accuracy （RMSE= 5.95%, R2=0.958） than the conventional LSMM method （RMSE=7.71%, R2=0.932）. Analysis with each vegeta- tion fraction, SVR model also show higher accuracy and stable ability showing excellent characteristics of non-linear spectrum unmixing. Hence, the high training ability can gained with the limitation sample amount. From the whole scene and each true value fraction, SVR method show better performance than that of LSMM due to the non-linear mixed pixel complexity which is the inevitable for LSMM. For SVR, the small amount train- ing sample can be used to represent the non-linear spectral space to determine the optimum hyperplane. Hence, the SVR is suitable to unmix the mixed pixels to solve the non-linear spectral problem. Furthermore, the SVR method can be applied for multi-endmember model from the actual remote sensing in the future to validate the performance of SVR unmixing model.