Caffe CNN-based classification of hyperspectral images on GPU

Deep learning techniques based on Convolutional Neural Networks (CNNs) are extensively used for the classification of hyperspectral images. These techniques present high computational cost. In this paper, a GPU (Graphics Processing Unit) implementation of a spatial-spectral supervised classification scheme based on CNNs and applied to remote sensing datasets is presented. In particular, two deep learning libraries, Caffe and CuDNN, are used and compared. In order to achieve an efficient GPU projection, different techniques and optimizations have been applied. The implemented scheme comprises Principal Component Analysis (PCA) to extract the main features, a patch extraction around each pixel to take the spatial information into account, one convolutional layer for processing the spectral information, and fully connected layers to perform the classification. To improve the initial GPU implementation accuracy, a second convolutional layer has been added. High speedups are obtained together with competitive classification accuracies.

     

基于GPU的空谱联合核稀疏表示高光谱分类并行优化

高光谱图像分类是遥感信息处理领域的热点问题,在核稀疏表示分类框架下,联合光谱信息和像元空间信息,空谱联合核稀疏表示高光谱图像分类能够取得较好的分类效果,但较高的计算复杂度及高光谱图像较大的数据量限制了其在实时性要求较高情况下的应用。基于GPU/CUDA架构,提出了一种空谱联合核稀疏表示高光谱分类的并行优化方法,设计访存优化策略对主机和设备端数据交互进行优化;充分利用GPU并行计算能力,加速分类过程中核矩阵的计算;采用依据GPU并行特性实现的矩阵运算,优化基于交替方向乘子法的分类模型求解过程。利用实际高光谱图像数据进行的实验,验证了该方法的有效性和高效性。     

多标签高光谱图像地物分类

目的 在高光谱地物分类中,混合像元在两个方面给单标签分类带来了负面影响：单类地物在混入异类地物后,其光谱特征会发生改变,失去独特性,使类内差异变大;多类地物在混合比例加深的情况下,光谱曲线会互相趋近,使类间差异变小。为了解决这一问题,本文将多标签技术运用在高光谱分类中。方法 基于高光谱特性,本文将欧氏距离与光谱角有机结合运用到基于类属属性的多标签学习LIFT（multi-label learning with label specific features）算法的类属属性构建中,形成了适合高光谱多标签的方法。基于标签地位的不相等,本文为多标签数据标注丰度最大标签,并在K最近邻KNN（k-nearest neighbor）算法中为丰度最大的标签设置比其余标签更大的权重,完成对最大丰度标签的分类。结果 在多标签分类与单标签分类的比较中,多标签表现更优,且多标签在precision指标上表现良好,高于单标签0.5% 1.5%。在与其余4种多标签方法的比较中,本文多标签方法在2个数据集上表现最优,在剩余1个数据集上表现次优。在最大丰度标签的分类上,本文方法表现优于单标签分类,在数据集Jasper Ridge上的总体分类精度提高0.2%,混合像元分类精度提高0.5%。结论 多标签分类技术应用在高光谱地物分类上是可行的,可以提升分类效果。本文方法根据高光谱数据的特性对LIFT方法进行了改造,在高光谱多标签分类上表现优异。高光谱地物的多标签分类中,每个像元多个标签的地位不同,在分类中可以通过设置不同权重体现该性质,提升分类精度。     

Tripartite-structure transformer for hyperspectral image classification

Hyperspectral images contain rich spatial and spectral information, which provides a strong basis for distinguishing different land-cover objects. Therefore, hyperspectral image (HSI) classification has been a hot research topic. With the advent of deep learning, convolutional neural networks (CNNs) have become a popular method for hyperspectral image classification. However, convolutional neural network (CNN) has strong local feature extraction ability but cannot deal with long-distance dependence well. Vision Transformer (ViT) is a recent development that can address this limitation, but it is not effective in extracting local features and has low computational efficiency. To overcome these drawbacks, we propose a hybrid classification network that combines the strengths of both CNN and ViT, names Spatial-Spectral Former(SSF). The shallow layer employs 3D convolution to extract local features and reduce data dimensions. The deep layer employs a spectral-spatial transformer module for global feature extraction and information enhancement in spectral and spatial dimensions. Our proposed model achieves promising results on widely used public HSI datasets compared to other deep learning methods, including CNN, ViT, and hybrid models.     

基于焦点损失的半监督高光谱图像分类

针对高光谱图像（HSI）训练数据获取困难的问题,采用了一种新的HSI半监督分类框架,该框架利用有限的标记数据和丰富的未标记数据来训练深度神经网络。同时,由于高光谱样本分布是不平衡的,导致不同样本分类难度存在巨大差异,采用原始交叉熵损失函数无法刻画这种分布特征,因而分类效果不理想。为了解决这个问题,在半监督分类框架中提出一种基于焦点损失的多分类目标函数。最后,考虑到HSI的空间信息对分类的影响,结合马尔可夫随机场（MRF）,利用样本空间特征进一步改善分类效果。在两个常用的HSI数据集上,将所提方法与多种典型算法进行了实验对比分析,实验结果表明所提方法能够产生优于其他对比方法的分类效果。     

样本优化选择的高光谱图像分类

目的 高光谱分类问题中,由于类内光谱特性存在差异性,导致常规的随机样本选择策略无法保证训练样本均匀覆盖样本空间。针对这一问题,提出基于类内再聚类的样本空间优化策略。同时为了进一步提高分类精度,针对低置信度分类结果,提出基于邻域高置信信息的修正策略。方法 采用FCM（fuzzy C-means）聚类算法对每类样本进行类内再聚类,在所聚的每个子类内选择适当样本。利用两个简单分类器SVM（support vector machine）和SRC（sparse representation-based classifier）,对分类结果进行一致性检测,确定高、低置信区域,对低置信区域,利用主成分图作为引导图对置信度图进行滤波,使得高置信信息向低置信区域传播,从而修正低置信区域分类结果。以上策略可以保证即便在较少的训练样本的情况下,也能够训练出较高的分类器,大幅度提高分类精度。结果 使用3组实验数据,根据样本比例设置两组实验与经典以及最新分类算法进行对比。实验结果表明,本文算法均取得很大改进,尤其在样本比例较小的实验中效果显著。在小比例（一般样本选取比例的十分之一）训练样本实验中,对于India Pines数据集,OA（overall accuracy）值高达90.48%;在Salinas数据集上能达到99.68%;同样,PaviaU数据集的OA值为98.54%。3组数据集的OA值均比其他算法高出4% 6%。结论 综上表明,本文算法通过样本空间优化策略选取有代表性、均衡性的样本,保证小比例样本下分类精度依然显著;基于邻域高置信信息的修正策略起到很好的优化效果。同时,本文算法适应多种数据集,具有很好的鲁棒性。     

一种高精度高光谱图像分类方案设计

为了有效改善高光谱图像数据分类的精确度,减少对大数目数据集的依赖,在原型空间特征提取方法的基础上,提出一种基于加权模糊C均值算法改进型原型空间特征提取方案。该方案通过加权模糊C均值算法对每个特征施加不同的权重,从而保证提取后的特征含有较高的有效信息量,从而达到减少训练数据集而不降低分类所需信息量的效果。实验结果表明,与业内公认的原型空间提取算法相比,该方案在相对较小的数据集下,其性能仍具有较为理想的稳定性,且具有相对较高的分类精度。     

Unsupervised real-time constrained linear discriminant analysis to hyperspectral image classification

We have proposed a constrained linear discriminant analysis (CLDA) approach for classifying the remotely sensed hyperspectral images. Its basic idea is to design an optimal linear transformation operator which can maximize the ratio of inter-class to intra-class distance while satisfying the constraint that the different class centers after transformation are aligned along different directions. Its major advantage over the traditional Fisher's linear discriminant analysis is that the classification can be achieved simultaneously with the transformation. The CLDA is a supervised approach, i.e., the class spectral signatures need to be known a priori. But, in practice, these informations may be difficult or even impossible to obtain. So in this paper we will extend the CLDA algorithm into an unsupervised version, where the class spectral signatures are to be directly generated from an unknown image scene. Computer simulation is used to evaluate how well the algorithm performs in terms of finding the pure signatures. We will also discuss how to implement the unsupervised CLDA algorithm in real-time for resolving the critical situations when the immediate data analysis results are required.     

Deep mixed precision for hyperspectral image classification

The Journal of Supercomputing - Hyperspectral images (HSIs) record scenes at different wavelength channels, providing detailed spatial and spectral information. How to storage and process this...     

基于 DenseNet-Attention 模型的高光谱图像分类

摘 要：针对高光谱图像标记样本量少,提取特征不充分以及提取到的特征不区分贡献度的问题,提出一个新型的 DenseNet-Attention 网络模型(DANet)。首先,该模型利用三维卷积核同步提取联合光谱空间特征,同时密集连接网络(DenseNet)的稠密连接块除了能够充分提取更加鲁棒的特征外,还减少了大量参数;其次,自注意力(self-attention)机制作为一个模块加入到稠密连接块中,可以使上层提取到的特征在进入下一层网络之前,经过该模块对其进行权重分配,使具有丰富的物类别信息的特征得到加强,进而区分特征的贡献度。网络模型以原始高光谱图像邻域块作为输入,无需任何预处理,是一个端对端学习的深度神经网络。在印第安松树林和帕维亚大学数据集上进行对比试验,网络模型的分类精度分别能够达到 99.43%和 99.99%,有效提高了高光谱图像分类精度。     

Effective naive Bayes nearest neighbor based image classification on GPU

Non-parametric classifier, Naive Bayes nearest neighbor, is designed with no training phase, and its performance outperforms many well-trained learning-based image classifiers. Unfortunately, despite its high accuracy, it suffers from great computational pressure from distance computations in space of local feature. This paper explores accelerating strategies from perspectives of both algorithm design and software development. Our approach integrates space decomposition capability of Product quantization (PQ) and parallel accelerating capability of underlying computational platform, Graphics processing unit (GPU). PQ is exploited to compress the indexed local features and prune the search space. GPU is used to ease most of computational pressure by processing the tasks in parallel. To achieve good parallel efficiency, a new sequential classification process is first designed and decomposed into independent components with high parallelism. Effective parallelization techniques are then presented to make use of computational resources. Parallel heap array is built to accelerate the process of feature quantization. Distance table lookup is built to speed up the process of feature search. Comparative experiments on UIUC-Sport dataset demonstrate that our integrated solution outperforms other implementations significantly on Core-quad Intel Core i7 950 CPU and GPU of NVIDIA Geforce GTX460. Scalability experiment on 80 million tiny images database shows that our approach still performs well when large-scale image database is explored.     

Unsupervised spectral sub-feature learning for hyperspectral image classification

Spectral pixel classification is one of the principal techniques used in hyperspectral image (HSI) analysis. In this article, we propose an unsupervised feature learning method for classification of hyperspectral images. The proposed method learns a dictionary of sub-feature basis representations from the spectral domain, which allows effective use of the correlated spectral data. The learned dictionary is then used in encoding convolutional samples from the hyperspectral input pixels to an expanded but sparse feature space. Expanded hyperspectral feature representations enable linear separation between object classes present in an image. To evaluate the proposed method, we performed experiments on several commonly used HSI data sets acquired at different locations and by different sensors. Our experimental results show that the proposed method outperforms other pixel-wise classification methods that make use of unsupervised feature extraction approaches. Additionally, even though our approach does not use any prior knowledge, or labelled training data to learn features, it yields either advantageous, or comparable, results in terms of classification accuracy with respect to recent semi-supervised methods.     

高光谱图像分类的自适应决策融合方法

目的 目前高光谱图像决策融合方法主要采用以多数票决（majority vote，MV）为代表的硬决策融合和以对数意见池（logarithmic opinion pool，LOGP）为代表的软决策融合策略。由于这些方法均使用统一的权重系数进行决策融合，没有对子分类器各自的分类性能进行评估而优化分配权重系数，势必会影响最终的分类精度。针对该问题，本文对多数票决和对数意见池融合策略进行了改进，提出了面向高光谱图像分类的自适应决策融合方法。方法 根据相关系数矩阵对高光谱图像进行波段分组，对每组波段进行空谱联合特征提取；利用高斯混合模型（Gaussian mixture model，GMM）或支持向量机（support vector machine，SVM）分类器对各组空谱联合特征进行分类；最后，采用本文研究的两种基于权重系数优化分配的自适应融合策略对子分类器的分类结果进行决策融合，使得分类精度低的波段组和异常值对最终分类结果的影响达到最小。结果 对两个公开的高光谱数据集分别采用多种特征和两种分类器组合进行实验验证。实验结果表明，在相同特征和分类器条件下，本文提出的自适应多数票决策融合策略（adjust majority vote，adjustMV）、自适应对数意见池决策融合策略（adjust logarithmic opinion pool，adjustLOGP）比传统的MV决策融合策略、LOGP决策融合策略对两个数据集的分类精度均有大幅度提高。Indian Pines数据集上，adjustMV算法的分类精度比相应的MV算法平均提高了1.2%，adjustLOGP算法的分类精度比相应的LOGP算法平均提高了7.38%；Pavia University数据集上，adjustMV算法的分类精度比相应的MV算法平均提高了2.1%，adjustLOGP算法的分类精度比相应的LOGP算法平均提高了4.5%。结论 本文提出的自适应权重决策融合策略为性能较优的子分类器（即对应于分类精度高的波段组）赋予较大的权重，降低了性能较差的子分类器与噪声波段对决策融合结果的影响，从而大幅度提高分类精度。所研究的决策融合策略的复杂度和计算成本均较低，在噪声环境中具有更强的鲁棒性，同时在一定程度上解决了高光谱图像分类应用中普遍存在的小样本问题。     

Discriminative graph convolution networks for hyperspectral image classification

Recently, the proposal of graph convolutional networks (GCN) has successfully implemented into hyperspectral image data representation and analysis. In spite of the great success, there are still several major challenges in hyperspectral image classification, including within-class diversity, and between-class similarity, which generally degenerate hyperspectral image classification performance. To address the problems, we propose a discriminative graph convolution networks (DGCN) for hyperspectral image classification. This method introduces the concepts of within-class scatter and between-class scatter, which respectively reflect the global geometric structure and discriminative information of the input space. The experimental results on the hyperspectral data sets show that the proposed method has good classification performance.     

Spectral-spatial K-Nearest Neighbor approach for hyperspectral image classification

Hyperspectral image (HSI) classification is a very active research topic in remote sensing and has numerous potential applications. This paper presents a simple but effective classification method based on spectral-spatial information and K-nearest neighbor (KNN). To be specific, we propose a spectral-spatial KNN (SSKNN) method to deal with the HSI classification problem, which effectively exploits the distances all neighboring pixels of a given test pixel and training samples. In the proposed SSKNN framework, a set-to-point distance is exploited based on least squares and a weighted KNN method is used to achieve stable performance. By using two standard HSI benchmark, we evaluate the proposed method by comparing it with eight competing methods. Both qualitative and quantitative results demonstrate our SSKNN method achieves better performance than other ones.     

基于支持向量机的高光谱影像分类研究

高光谱遥感技术,将反映目标辐射属性的光谱信息与反映目标空间几何关系的图像信息有机地结合在一起.高光谱影像丰富的光谱信息使其较全色遥感、多光谱遥感能够更好的进行地面目标的分类识别.本文综合利用支持向量机分类的若干关键技术,包括序列最小优化训练算法,多类支持向量机构造方法、核函数及其参数选择的交叉验证"网格搜索",给出了高光谱影像分类流程,进行了遥感数据试验分析.     

A new approach to mixed pixel classification of hyperspectral imagery based on extended morphological profiles

This paper presents a new approach to the analysis of hyperspectral images, a new class of image data that is mainly used in remote sensing applications. The method is based on the generalization of concepts from mathematical morphology to multi-channel imagery. A new vector organization scheme is described, and fundamental morphological vector operations are defined by extension. Theoretical definitions of extended morphological operations are used in the formal definition of the concept of extended morphological profile, which is used for multi-scale analysis of hyperspectral data. This approach is particularly well suited for the analysis of image scenes where most of the pixels collected by the sensor are characterized by their mixed nature, i.e. they are formed by a combination of multiple underlying responses produced by spectrally distinct materials. Experimental results demonstrate the applicability of the proposed technique in mixed pixel analysis of simulated and real hyperspectral data collected by the NASA/Jet Propulsion Laboratory Airborne Visible/Infrared Imaging Spectrometer and the DLR Digital Airborne (DAIS 7915) and Reflective Optics System Imaging Spectrometers. The proposed method works effectively in the presence of noise and low spatial resolution. A quantitative and comparative performance study with regards to other standard hyperspectral analysis methodologies reveals that the combined utilization of spatial and spectral information in the proposed technique produces classification results which are superior to those found by using the spectral information alone.     

双卷积池化结构的3D-CNN高光谱遥感影像分类方法

目的 高光谱遥感影像数据包含丰富的空间和光谱信息,但由于信号的高维特性、信息冗余、多种不确定性和地表覆盖的同物异谱及同谱异物现象,导致高光谱数据结构呈高度非线性。3D-CNN（3D convolutional neural network）能够利用高光谱遥感影像数据立方体的特性,实现光谱和空间信息融合,提取影像分类中重要的有判别力的特征。为此,提出了基于双卷积池化结构的3D-CNN高光谱遥感影像分类方法。方法 双卷积池化结构包括两个卷积层、两个BN（batch normalization）层和一个池化层,既考虑到高光谱遥感影像标签数据缺乏的问题,也考虑到高光谱影像高维特性和模型深度之间的平衡问题,模型充分利用空谱联合提供的语义信息,有利于提取小样本和高维特性的高光谱影像特征。基于双卷积池化结构的3D-CNN网络将没有经过特征处理的3D遥感影像作为输入数据,产生的深度学习分类器模型以端到端的方式训练,不需要做复杂的预处理,此外模型使用了BN和Dropout等正则化策略以避免过拟合现象。结果 实验对比了SVM（support vector machine）、SAE（stack autoencoder）以及目前主流的CNN方法,该模型在Indian Pines和Pavia University数据集上最高分别取得了99.65%和99.82%的总体分类精度,有效提高了高光谱遥感影像地物分类精度。结论 讨论了双卷积池化结构的数目、正则化策略、高光谱首层卷积的光谱采样步长、卷积核大小、相邻像素块大小和学习率等6个因素对实验结果的影响,本文提出的双卷积池化结构可以根据数据集特点进行组合复用,与其他深度学习模型相比,需要更少的参数,计算效率更高。     

高斯过程及其在高光谱图像分类中的应用

高光谱遥感图像分类是高光谱成像信息处理的研究热点,高光谱成像的内在特点对于分类器设计具有直接影响.高斯过程是近年来发展迅速的一种新的机器学习方法,具备容易实现、超参数可自适应获取以及预测输出具有概率意义等优点,比较适合于处理图像分类问题.首先对高斯过程的基本概念及其主要的分类算法进行了简要介绍,然后在对高光谱图像分类的特点和高光谱图像分类的研究现状的分析基础上,讨论了基于高斯过程的高光谱图像分类的基本思想,提出了基于空间约束的高斯过程分类和基于半监督高斯过程分类等适合高光谱图像分类的新方法.最后对基于高斯过程的高光谱图像分类研究的发展趋势进行了展望.