提出一种新的基于高斯误差函数(Gaussian error function, Erf)作为激活函数的BP神经网络(Erf-BP),并用于林区TM影像进行混合像元分解。研究表明:Erf-BP模型的精度高于线性无约束最小二乘法模型及最大似然法。通过在高分辨率影像上选取验证样区精度检验得出:1) 各端元总分解精度为89.2%,RMSE比线性无约束最小二乘法模型降低了近39%; 2) 该方法能够较高精度地提取森林遥感信息,精度达到86%,RMSE比线性无约束最小二乘法模型降低了近40.6%。将3种不同方法估计的整个研究区各端元面积百分比与森林资源二类调查数据作对比得出:Erf-BP模型精度略高于最大似然法,RMSE分别为4.18%和7.90%,两者精度明显高于线性无约束最小二乘法模型(RMSE=18.75%)。Erf-BP算法能够较高精度地对TM影像进行混合像元分解,尤其在森林信息提取上,为基于混合像元分解提取不同森林类型甚至树种遥感信息提供一种可行的方法。
A new approach based on Gaussian error function back-propagation(Erf-BP)neural network was developed to analyze mixture pixels and was applied in forest area in Anji County, Zhejiang Province. The study results showed that Erf-BP model was superior to unconstrained linear spectral mixture analysis model and maximum likelihood method. Through collecting sample plots from the high-resolution satellite imagery to evaluate accuracy, the results showed: the total accuracy yielded 89.2% for the Erf-BP model, and RMSE was approximately 39% lower than unconstrained linear spectral mixture analysis model.For forest information extraction, the accuracy yielded 86% for the Erf-BP model, and RMSE was approximately 40.6% lower than unconstrained linear spectral mixture analysis model. At the same time, compared the area percent of each endmember estimated from the three methods with forest resource inventory data, the results showed the accuracy of Erf-BP model (RMSE=4.18%) was slightly higher than maximum likelihood method (RMSE=7.90%) and obviously higher than unconstrained linear spectral mixture analysis model (RMSE=18.75%). Erf-BP model was a feasible method to extract remote information of different forest types, even of different tree species.