深度置信网络环境下高光谱数据降维方法仿真
|
摘要
为解决当前高光谱数据降维方法存在的降维效率低、丢包现象严重等问题,提出基于Manifold的深度置信网络环境下高光谱数据降维方法。根据高光谱数据定义域和值域的归一化,通过超平面最小策略实现高光谱数据去噪问题的转换,即将高光谱数据去噪问题转为能量的最小化问题。获取与能量最小化问题相对应的非线性Euler-Lagrange方程,利用其迭代形式完成对其唯一解的求取。将权重系数引入方程求解中,实现高光谱数据去噪。计算两个高光谱数据样本邻近区域之间的距离,确定各高光谱数据点与其邻近点于一个Manifold线性区域中共存,并得到高维高光谱数据全局结构信息。据此计算数据空间样本权值,利用该权值的平移和旋转以及缩放特性,获得高维高光谱数据降维结果。仿真表明,上述方法运行下平均丢包率约为0.4%,数据降维效率较高。所提方法具有较强的降维性能,相比当前相关方法更具可借鉴性。
In current methods, the efficiency of dimension reduction is low and the phenomenon of packet loss is serious. Therefore, this article focuses on a method of dimension reduction for hyperspectral data in deep confidence network environment based on Manifold. According to the normalization of domain and range of hyperspectral data, the hyperplane minimum strategy was used to transform the problem about noise reduction of hyperspectral data. In other words, the problem about noise reduction of hyperspectral data was converted into the problem of energy minimization. Then, nonlinear Euler-Lagrange equation corresponding to the energy minimization problem was obtained. In addition, its iterative form was used to find the unique solution. Moreover, the weight coefficient was introduced into the equation to achieve noise reduction of hyperspectral data. Meanwhile, the distance between the adjacent regions of two hyperspectral data samples and each hyperspectral data point and its neighbor point were coexisted in a Manifold linear region. Thus, the global structure information of high-dimensional hyperspectral data was obtained. On this basis, the data space sample weight was calculated. Finally, the characteristics of weight such as translation, rotation and scaling were used to obtain the dimension reduction result of high-dimensional hyperspectral data. Simulation shows that the average packet loss rate with above method is about 0.4% and the efficiency of dimension reduction is high. Meanwhile, the proposed method has strong performance of dimension reduction, which has some references.
In current methods, the efficiency of dimension reduction is low and the phenomenon of packet loss is serious. Therefore, this article focuses on a method of dimension reduction for hyperspectral data in deep confidence network environment based on Manifold. According to the normalization of domain and range of hyperspectral data, the hyperplane minimum strategy was used to transform the problem about noise reduction of hyperspectral data. In other words, the problem about noise reduction of hyperspectral data was converted into the problem of energy minimization. Then, nonlinear Euler-Lagrange equation corresponding to the energy minimization problem was obtained. In addition, its iterative form was used to find the unique solution. Moreover, the weight coefficient was introduced into the equation to achieve noise reduction of hyperspectral data. Meanwhile, the distance between the adjacent regions of two hyperspectral data samples and each hyperspectral data point and its neighbor point were coexisted in a Manifold linear region. Thus, the global structure information of high-dimensional hyperspectral data was obtained. On this basis, the data space sample weight was calculated. Finally, the characteristics of weight such as translation, rotation and scaling were used to obtain the dimension reduction result of high-dimensional hyperspectral data. Simulation shows that the average packet loss rate with above method is about 0.4% and the efficiency of dimension reduction is high. Meanwhile, the proposed method has strong performance of dimension reduction, which has some references.
引文
[1] 方民权,张卫民,周海芳.集成众核上快速独立成分分析降维并行算法[J].计算机研究与发展,2016,53(5):1136-1146.
[2] 戴晓爱,等.空间数据压缩的高光谱降维技术比较[J].遥感信息,2017,32(2):107-112.
[3] 刘韵佳,赵荣珍,王雪冬.基于Schur分解和正交邻域保持嵌入算法的故障数据集降维方法[J].中国机械工程,2017,28(21):2552-2556.
[4] 史华.基于复杂网络的建筑物强震下抗毁性估计模型[J].地震工程学报,2017,39(6):1024-1028.
[5] 鲁锦涛,马丽.基于流形对齐的高光谱遥感图像降维和分类算法[J].国土资源遥感,2017,29(1):104-109.
[6] 高红民,等.神经网络敏感性分析的高光谱遥感影像降维与分类方法[J].电子与信息学报,2016,38(11):2715-2723.
[7] 陈鹏飞,等.等距映射和局部线性嵌入算法集成的转子故障数据集降维方法[J].振动与冲击,2017,36(6):45-50.
[8] 翁谦,等.应用谱回归和图正则最小二乘回归的数据降维[J].计算机工程与应用,2017,53(5):81-84.
[9] 王文俊.基于稀疏类别保留投影的基因表达数据降维方法[J].电子学报,2016,44(4):873-877.
[10] 顾爱华.云计算网络中高维数据标准化处理优化仿真[J].计算机仿真,2017,34(3):317-320.
[2] 戴晓爱,等.空间数据压缩的高光谱降维技术比较[J].遥感信息,2017,32(2):107-112.
[3] 刘韵佳,赵荣珍,王雪冬.基于Schur分解和正交邻域保持嵌入算法的故障数据集降维方法[J].中国机械工程,2017,28(21):2552-2556.
[4] 史华.基于复杂网络的建筑物强震下抗毁性估计模型[J].地震工程学报,2017,39(6):1024-1028.
[5] 鲁锦涛,马丽.基于流形对齐的高光谱遥感图像降维和分类算法[J].国土资源遥感,2017,29(1):104-109.
[6] 高红民,等.神经网络敏感性分析的高光谱遥感影像降维与分类方法[J].电子与信息学报,2016,38(11):2715-2723.
[7] 陈鹏飞,等.等距映射和局部线性嵌入算法集成的转子故障数据集降维方法[J].振动与冲击,2017,36(6):45-50.
[8] 翁谦,等.应用谱回归和图正则最小二乘回归的数据降维[J].计算机工程与应用,2017,53(5):81-84.
[9] 王文俊.基于稀疏类别保留投影的基因表达数据降维方法[J].电子学报,2016,44(4):873-877.
[10] 顾爱华.云计算网络中高维数据标准化处理优化仿真[J].计算机仿真,2017,34(3):317-320.