基于卷积神经网的高光谱数据特征提取及分类技术研究

发布时间：2017-12-28 00:21

  本文关键词：基于卷积神经网的高光谱数据特征提取及分类技术研究　出处：《哈尔滨工业大学》2016年硕士论文　论文类型：学位论文

  更多相关文章： 高光谱图像 深度学习 卷积神经网络 特征提取 分类

【摘要】：基于高光谱数据的特征提取及分类技术一直是遥感领域研究的热点问题之一,而现有的特征提取方法主要针对地物某一方面的特性,利用线性或非线性的方程人为地设计或指定提取的特征,这种人工选取特征的过程往往需要专业的知识和经验,并且需要花费大量的时间,然而提取的特征并不能充分表达高光谱数据复杂的内部结构和空谱信息。对于深度学习来说,它可以让计算机自动地学习有利于任务需要的特征,并将该过程融入模型训练的一部分,从而有助于进一步提高分类识别精度。本篇论文从高光谱数据的特点入手,结合基于深度学习的卷积神经网络模型,利用多个卷积层和池化层从高光谱数据中提取对多种变形具有高度不变性的非线性特征,进而实现高光谱数据的地物分类。本文的主要研究内容及成果包括以下几个方面:首先,针对高光谱遥感数据图谱合一的特点,探究深层卷积网络对高光谱数据特征提取及分类的适用性。高光谱数据在获取拍摄面的空间信息时,可以获得每一个像素的连续光谱曲线,这使得高光谱数据拥有较高的维度和较大的数据量,而深度学习的模型正适用于该数据的特点。因此本文使用高光谱数据的光谱信息、空间信息和空谱联合信息,分别构造基于一维、二维和三维卷积核的深层卷积神经网络,实现了特征分级式表达,并将提取的特征引入高光谱数据的地物分类中,得到优于其他特征提取及分类方法的结果。其次,针对数据高维度与有限训练样本的不均衡问题,本文分别在一维卷积模型中引入L2正则项修改原始代价函数,在二维和三维模型中加入Dropout层稀疏每层网络的激活单位,来避免建模过程中过拟合现象的发生。并利用非饱和的非线性函数Re LU代替原有的Sigmoid激活函数,大大提高模型的收敛速度,降低了模型的复杂度。最后,在较少的训练样本下,基于光谱信息的一维深层卷积模型并不能提供稳定的分类结果,为了进一步提高模型的分类性能,本文提出了基于随机特性选择的深度卷积神经网络集成模型。从两组数据的实验结果表明,与其他分类方法的分类精度相比,该方法是一个较有竞争力的解决方案。
[Abstract]:Based on the feature extraction and classification of hyperspectral data has been one of the hot issues in the field of remote sensing research, and the existing methods of feature extraction for the main features of a certain objects, using linear or nonlinear equations artificially specified design or feature extraction, the artificial feature selection process often requires professional knowledge and experience. And the need to spend a lot of time, however, the extracted features and can not fully express the internal structure information of hyperspectral data and complicated spatial spectrum. For deep learning, it allows the computer to automatically learn features that are beneficial to the needs of tasks, and integrate the process into part of model training, which is helpful to further improve the accuracy of classification and recognition. This paper starts from the characteristic of hyperspectral data, based on convolutional neural network model based on deep learning, using a multi layer and pool layer extraction volume is highly nonlinear deformation characteristics of multiple invariance from hyperspectral data, so as to realize the classification of hyperspectral data. The main contents and achievements of this paper include the following aspects: first, aiming at the characteristics of hyperspectral remote sensing data, we explore the applicability of deep convolution network to hyperspectral data feature extraction and classification. Hyperspectral data can get continuous spectral curves of each pixel when acquiring the spatial information of the shooting surface, which makes the hyperspectral data have higher dimension and larger data volume, while the deep learning model is suitable for the characteristics of the data. The spectral information and spatial information and it is the use of hyperspectral data with spectral information, we construct one-dimensional, two-dimensional and three-dimensional convolution kernel deep convolutional neural network based on the expression characteristics of hierarchical, and feature extraction of hyperspectral data into object classification, feature extraction and classification is better than that by other methods the results of. Secondly, aiming at the problem of unbalanced data in high dimension and limited training samples, this paper in the convolution model introduced by L2 regularization to modify the original cost function, activation of Dropout units into each layer of the network layer is sparse in 2D and 3D model, to avoid the overfitting phenomenon in the process of modeling. And using the unsaturated nonlinear function Re LU instead of the original Sigmoid activation function, the convergence speed of the model is greatly improved and the complexity of the model is reduced. Finally, under a few training samples, the one-dimensional deep convolution model based on spectral information can not provide stable classification results. In order to further improve the classification performance of the model, a deep convolution neural network ensemble model based on random feature selection is proposed in this paper. The experimental results from two sets of data show that this method is a more competitive solution compared with the classification accuracy of other classification methods.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP751;TP183
，

本文编号：1343820