基于分布式信源编码的高光谱图像无损压缩研究进展

有效的高光谱图像压缩技术已经成为航天高光谱遥感领域研究的焦点之一。对基于分布式信源编码（Distributed Source Coding,DSC）的高光谱图像压缩技术研究进展进行了综述。首先介绍了DSC的理论基础、实现方式及其在高光谱图像无损压缩应用中的优势;然后总结了基于DSC的高光谱图像无损压缩研究进展,在此基础上给出了一种基于多波段预测的高光谱图像分布式无损压缩算法,实验结果表明,该算法具有较低的编码复杂度,其无损压缩性能优于现有的分布式无损压缩算法;最后指出了DSC在高光谱图像压缩中需要进一步研究的问题。     

基于三维卷积残差网络的无人机高光谱岩性分类

岩性识别和分类是地质学、资源勘查等不可或缺的环节,高光谱遥感的兴起为岩性识别提供新的思路。利用机器学习挖掘岩石高光谱图像中的信息从而准确识别岩性,这具有重要的应用价值。目前用机器学习的方法实现岩石的高光谱影像分类研究中,缺少对空间和光谱信息的充分利用,因此本文使用了一种加入注意力机制的三维卷积残差网络结构,能够有效提取岩石高光谱图像的空间、光谱特征以及空谱联合特征。本实验利用无人机搭载高光谱传感器采集了10种不同类型的岩石样本影像,应用该算法对岩石高光谱图像进行分类。实验结果表明：该算法与传统机器学习算法SVM、RF和深度学习算法ResNet、3D CNN和SSRN相比具有更高的精度。     

KDTGAN：基于Transformer-GAN和知识蒸馏的高光谱目标检测

高光谱目标检测在地球观测中至关重要,被广泛应用于军事和民用领域。然而,由于高光谱图像的背景复杂性和目标样本的有限性,该任务面临较大的挑战。本文首先采用CEM（约束能量最小化）粗检测方法提取背景数据。随之,引入了一种新的知识蒸馏模型,即KDTGAN（通过Transformer-GAN实现）。教师模型的生成器采用了Transformer编码器的结构,并结合多尺度数据融合的方法,能够准确地学习背景分布,进而通过重构背景信息实现目标检测。为了克服GAN（生成对抗网络）训练不稳定的挑战,特别是纯背景数据的稀缺性,本文提出了一种新的损失算法,以减小可疑目标样本对模型性能的负面影响。为了降低模型的计算负担,本文引入知识蒸馏,并设计新的蒸馏损失对学生模型加以约束,使模型轻量化的同时提高学生模型检测精度。实验结果表明：KDTGAN相较于当前检测方法表现更优,具有更高的检测精度和鲁棒性。     

Novel method for reconstruction of hyperspectral resolution images from multispectral data for complex coastal and inland waters

Hyperspectral resolution image products of a synthetic sensor featuring the high spatial resolution of the space-borne sensor can offer cost-effective means for enhancing our current capabilities in terms of providing an array of images in lieu of designing an expensive system for image acquisition, which can serve the expanding needs of the scientific and user communities for various critical water color applications. Despite several studies on enhancing the capability of land remote sensing sensors, full spectrum reconstruction of water color images with varying spectral bands is hampered by the lack of methods and accurate atmospheric correction procedures. In the present work, a novel method is developed for reconstruction of hyperspectral resolution images from high spatial-resolution Sentinel 2 Multispectral Instrument (MSI) data representative of many complex waters in coastal and inland zones. This method uses a deep neural network (DNN) with multiple blocks of deconvolution and dense layers. The spectral reconstruction of hyperspectral resolution images from multispectral data was based on rigorous training data from the atmospherically-corrected and validated HICO normalized water-leaving radiance products (with spectral resolution 438-868 nm sampled at 5.7 nm) of diverse water types. The generalizability and versatility of the DNN method was tested and evaluated systematically by means of various qualitative and quantitative analyses using concurrent space-borne (MSI and HICO) and in-situ measurements from different regional waters. Reconstructed hyperspectral resolution radiances obtained from the MSI images closely matched with independent HICO and MSI measurements within the desired accuracy. Successful reconstruction and validation of the hyperspectral radiances indicate that the proposed state-of-the-art method provides possible future directions for enhancing our current capabilities of space-borne sensors for various research purposes and societal applications at local, regional and global scales.     

Fractal signatures for multiscale processing of hyperspectral image data

A multiscale approach to hyperspectral image data analysis using fractal signatures was proposed and implemented in the Interactive Data Language (IDL). For 2-D hyperspectral curves, fractal signature measures the changes in curve length with changing scale. Using NASA’s Earth Observing-1 (EO-1) Hyperion image from a study area near Denton, Texas, USA, the capabilities of fractal signatures in discriminating different land cover types were presented in three different ways: (1) fractal signature curves, (2) distances between fractal signatures, and (3) fractal signature images. The asymmetry in length measurement was found to be effective in handling hyperspectral curves obtained from Hyperion radiance data. The contribution of fractal signature images was shown through comparison of image classification results. The results from the Hyperion radiance data suggest that fractal signatures at certain scales can reveal important differences in land cover types.     

Unmixing method for hyperspectral data based on sub-space method with learning process

An unmixing method for hyperspectral Earth observation satellite imagery data is proposed. It is based on a sub-space method with learning process. The proposed method utilizes a sub-space for feature space during unmixing. It is used to be done in a feature space which consists of spectral bands of observation vectors. As the results from the experiments with airborne based hyperspectral imagery data, AVIRIS, it is found that the proposed unmixing is superior to the other existing method in terms of decomposition accuracy and the process time required for the decompositions.     

The PRISMA payload optomechanical design, a high performance instrument for a new hyperspectral mission

PRISMA (PRecursore IperSpettrale della Missione Applicativa) hyperspectral instrument is an advanced hyperspectral sensor including a panchromatic camera at medium resolution. The instrument is the focus of the new Earth observation mission that a consortium of Italian companies has started developing under contract of Italian Space Agency. Key features of the instrument are the very high requirement for signal-to-noise and the high quality of data that have to be provided. To meet these demanding figures the optical system has been based on a high transmittance optical system, including a single mirror telescope and two prism spectrometers based on an innovative concept to minimize number of optical elements, while high performance detectors have been chosen for the photon detection. To provide the required data quality for the entire mission lifetime an accurate calibration unit (radiometric and spectral) will be included in the instrument optomechanical assembly. The thermo-mechanical design of the instrument is based on innovative concepts, considering that the use of prism spectrometers implies a tight control of temperature variations to guarantee the stability of all instrument features once in orbit. The presented paper describes the concepts and design principle of the optomechanical assembly of the instrument, at the present status of development.     

Classification of hyperspectral remote-sensing data with primal SVM for small-sized training dataset problem

With recent technological advances in remote sensing, very high-dimensional (hyperspectral) data are available for a better discrimination among different complex land-cover classes having similar spectral signatures. However, this large number of bands makes very complex the task of automatic data analysis. In the real application, it is difficult and expensive for the expert to acquire enough training samples to learn a classifier. This results in a classification problem with small-size training sample set. Recently, a regularization-based algorithm is usually proposed to handle such problem, such as Support Vector Machine (SVM), which usually are implemented in the dual form with Lagrange theory. However, it can be solved directly in primal formulation. In this paper, we introduces an alternative implementation technique for SVM to address the classification problem with small-size training sample set. It has been empirically proven that the effectiveness of the introduced implementation technique which has been evaluated by benchmark datasets.     

基于混合Contourlet变换和主成分变换的高光谱图像融合

针对高光谱图像波段多、存在噪声干扰等特点,提出了一种混合Contourlet和主成 分 分析（principal component analysis, 简称PCA）变换的高光谱图像融合方法。首先将多 个波段的高光谱图像进行Contourlet变换,得到系列多尺度、方向各异的子带系数,然后利 用PCA变换对各子带系数分别进行自适应融合处理。实验结果表明该算法可以有效地进行高 光谱图像融合,消除噪声干扰,获得比直接应用Contourlet变换和PCA变换更好的融合效果 。〖JP〗