可变神经网络结构下的遥感影像光谱分解方法  

Spectral Unmixing Method of Remote Sensing Images in Variable Architecture of Neural Network

作　　者：李熙[1] 石长民 李畅[3] 陈锋锐[4] 田礼乔[1] 

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,武汉430079 [2]三亚市国土环境资源信息中心,海南三亚572000 [3]华中师范大学城市与环境科学学院,武汉430079 [4]河南大学环境与规划学院,河南开封475000

出　　处：《计算机工程》2012年第9期1-3,共3页Computer Engineering

基　　金：国家自然科学基金资助项目(41101413);高等学校博士学科点专项科研基金资助项目(20110141120073);中央高校基本科研业务费专项基金资助项目(904275839)

摘　　要：多层感知神经网络(MLP)是主流的非线性分解方法,但是目前缺乏有效方法处理MLP分解结果中的丰度负值问题。为此,提出一种可变神经网络结构的方法,逐步去除负值丰度对应的端元,并调整相应的网络结构使之针对剩余的端元进行分解。通过武汉地区模拟TM遥感影像实验可以发现,该方法与传统MLP方法以及线性光谱分解方法的平均误差分别为0.077 7、0.081 9、0.094 3,说明该方法的分解精度高于其他2种分解方法,能克服丰度负值问题。Spectral unmixing of remote sensing images is a hotspot in remote sensing field, and Multilayer Perception（MLP） neural network is a common nonlinear spectral unmixing algorithm. However, currently there is no effective way to deal with the negative abundances derived by the network. To solve this problem, a MLP neural network with variable architecture is proposed. By discarding endmembers with negative abundances, the MLP architecture is modified to unmix the rest endmembers, so a remote sensing image is finally unmixed. An experiment using a simulated TM image shows that the average errors of the proposed method, conventional MLP method and linear spectral unmixing model are 0.077 7, 0.081 9 and 0.094 3 respectively, thus the proposed method outperforms the other two. Therefore, the proposed method can overcome the negative abundance problem effectively.

关 键 词：遥感 混合像元 神经网络 多层感知网络 非负约束 非线性光谱分解模型 

分 类 号：TN911.73[电子电信—通信与信息系统]