Using low‐spectral‐resolution images to acquire simulated hyperspectral images

We propose a method to acquire simulated hyperspectral images using low‐spectral‐resolution images. Hyperspectral images provide more spectral information than low‐spectral‐resolution images, because of the additional spectral bands used for data acquisition in hyperspectral imaging. Unfortunately, original hyperspectral images are more expensive and more difficult to acquire. However, some research questions require an abundance of spectral information for ground monitoring, which original hyperspectral images can easily provide. Hence, we need to propose a method to acquire simulated hyperspectral images, when original hyperspectral images are especially necessary. Since low‐spectral‐resolution images are readily available and cheaper, we develop a method to acquire simulated hyperspectral images using low‐spectral‐resolution images. With simulated hyperspectral images, we can acquire more ‘hidden’ information from low‐spectral‐resolution images. Our method uses the principles of pixel‐mixing to understand the compositional relationship of spectrum data to an image pixel, and to simulate radiation transmission processes. To this end, we use previously obtained data (i.e. spectrum library) and the sorting data of objects that are derived from a low‐spectral‐resolution image. Using the simulation of radiation transmission processes and these different data, we acquire simulated hyperspectral images. In addition, previous analyses of simulated remotely sensed images do not use quantitative statistical measures, but use qualitative methods, describing simulated images by sight. Here, we quantitatively assess our simulation by comparing the correlation coefficients of simulated images and real images. Finally, we use simulated hyperspectral images, real Hyperion images, and their corresponding ALI images to generate several classification images. The classification results demonstrate that simulated hyperspectral data contain additional information not available in the multispectral data. We find that our method can acquire simulated hyperspectral images quickly.     

遥感图像飞机目标分类的卷积神经网络方法

目的 遥感图像飞机目标分类,利用可见光遥感图像对飞机类型进行有效区分,对提供军事作战信息有重要意义。针对该问题,目前存在一些传统机器学习方法,但这些方法需人工提取特征,且难以适应真实遥感图像的复杂背景。近年来,深度卷积神经网络方法兴起,网络能自动学习图像特征且泛化能力强,在计算机视觉各领域应用广泛。但深度卷积神经网络在遥感图像飞机分类问题上应用少见。本文旨在将深度卷积神经网络应用于遥感图像飞机目标分类问题。方法 在缺乏公开数据集的情况下,收集了真实可见光遥感图像中的8种飞机数据,按大致4∶1的比例分为训练集和测试集,并对训练集进行合理扩充。然后针对遥感图像与飞机分类的特殊性,结合深度学习卷积神经网络相关理论,有的放矢地设计了一个5层卷积神经网络。结果 首先,在逐步扩充的训练集上分别训练该卷积神经网络,并分别用同一测试集进行测试,实验表明训练集扩充有利于网络训练,测试准确率从72.4%提升至97.2%。在扩充后训练集上,分别对经典传统机器学习方法、经典卷积神经网络LeNet-5和本文设计的卷积神经网络进行训练,并在同一测试集上测试,实验表明该卷积神经网络的分类准确率高于其他两种方法,最终能在测试集上达到97.2%的准确率,其余两者准确率分别为82.3%、88.7%。结论 在少见使用深度卷积神经网络的遥感图像飞机目标分类问题上,本文设计了一个5层卷积神经网络加以应用。实验结果表明,该网络能适应图像场景,自动学习特征,分类效果良好。     

Landsat approaches to map agro-pastoral farming in the wetlands of southern Sudan

Machine-learning algorithms (MLA) are coming of age within satellite remote sensing (SRS). This study compares the performance of a number of MLAs with more traditional indices and algorithms to map annual agro-pastoralist farming activity in southern Sudan. Two Landsat images from the early dry season 2014 and 2015 were analysed thoroughly and evaluated by interpretation of farming cover from very high resolution (VHR) images on Google Earth (GE). Traditional SRS indices based upon red and near infrared (NIR) bands used for monitoring rangelands did not perform well for the wet rangeland conditions compared to the use of blue and shortwave infrared (SWIR) bands. The species distribution model programme, MaxEnt, was used to produce a continuous farming activity indices using only Landsat-derived variables. Compared to other SRS classification approaches, maximum entropy (MaxEnt) showed the best overall performance to map farming activity followed by classification tree analysis (CTA). Overall mapping agreement >95.0% was reached for most methodologies, with MaxEnt showing very high mapping agreement (≥98.5%) for both years. When the result of MaxEnt’s good performance is put together in a 2014–15 or a 1999–2002 change detection scenario, it corroborates ground reports on massive human abuses that have taken place in Unity state of southern Sudan.     

基于密集卷积神经网络的遥感飞机识别

传统的飞机识别方法受模糊、遮挡、噪声以及光照等多种因素的干扰时会降低识别率,且卷积神经网络主要依赖局部特征,却丢失了轮廓特征等重要的全局结构化特征,从而导致算法对于受干扰飞机图像识别效果不佳。因此,基于密集卷积神经网络提出一种结合局部与全局特征的联合监督识别方法,以密集卷积神经网络为基础得到图像特征,通过结合局部特征（卷积神经网络特征）与全局特征（方向梯度直方图特征）进行分类,分类器目标函数使用softmax损失和中心损失联合监督方法。实验结果表明,局部特征与全局特征的结合使算法更加智能化,且损失函数联合监督方法能够实现图像深层特征的类内聚合、类间分散,该算法能有效解决卷积神经网络对受到多种干扰的遥感图像识别率低的问题。     

Land Cover Classification for Different Spatial Resolution Images from CNN

Based on convolutional neural networks and five different spatial resolution remote sensing images, the land use/land cover classification study was carried out on a small area in the eastern part of Xining City, aiming at exploring the differences of image classification by CNN with different spatial resolutions and CNN’s ability to extract different features. In order to improve the selection efficiency of the samples, a window sliding method was introduced to assist the samples selection. The research shows that the overall classification accuracy of the five different spatial resolution images is above 89%, the Kappa coefficient is above 0.86. The result further shows that within the resolution scale the higher the resolution, the performance of the CNN classification results for the details is better, and can maintain high classification accuracy, indicating that CNN is more suitable for high spatial resolution images; at the same time, the image spatial resolution is too high, the ground objects exhibit high intra-class variability and low inter-class variability, the classification accuracy tends to decrease. In comparison, CNN has the best classification effect on SPOT 6 images in this study, and window sliding is an effective sample-assisted selection method. This research has certain reference significance for similar research in the future.     

基于CNN的不同空间分辨率影像土地覆被分类研究

基于卷积神经网络（Convolutional Neural Networks, CNN）和5种不同空间分辨率的遥感影像,对西宁市东部一区域开展土地覆被分类研究,旨在探索CNN在不同空间分辨率下进行影像分类的差异性和对不同地物的提取能力。为提高样本的选择效率,引入了窗口滑动方法进行辅助选样。研究表明5种不同空间分辨率影像的总体分类精度均达89%以上,Kappa系数达0.86以上,分类精度较高。在所涉及的分辨率尺度范围内,空间分辨率越高,CNN分类结果越精细,并能保持较高的分类精度,表明CNN更适合高空间分辨率影像分类;但同时影像空间分辨率越高,地物表现出较高的类内变异性和低类间差异性,分类精度有降低的趋势。相比较而言,SPOT 6影像的分类精度最高,同时窗口滑动是一种有效的样本辅助选择方法。研究对今后同类工作具有一定的借鉴意义。     

Pattern recognition experiments in the mandala/cosine domain

The problem of recognition of objects in images is investigated from the simultaneous viewpoints of image bandwidth compression and automatic target recognition. A scenario is suggested in which recognition is implemented on features in the block cosine transform domain which is useful for data compression as well. While most image frames would be processed by the automatic recognition algorithms in the compressed domain without need for image reconstruction, this still allows for visual image classification of targets with poor recognition rates (by human viewing at the receiving terminal). It has been found that the Mandala sorting of the block cosine domain results in a more effective domain for selecting target identification parameters. Useful features from this Mandala/cosine domain are developed based upon correlation parameters and homogeneity measures which appear to successfully discriminate between natural and man-made objects. The Bhattacharyya feature discriminator is used to provide a 10:1 compression of the feature space for implementation of simple statistical decision surfaces (Gaussian and minimum distance classification). Imagery sensed in the visible spectra with a resolution of approximately 5-10 ft is used to illustrate the success of the technique on targets such as ships to be separated from clouds. A data set of 38 images is used for experimental verification with typical classification results ranging from the high 80's to low 90 percentile regions depending on the options choosen.     

Speckle reduction for the forest mapping analysis of multi-temporal Radarsat-1 images

As the number of satellite-borne synthetic aperture radar (SAR) systems increases, both the availability and the length of multi-temporal (MT) sequences of SAR images have also increased. Previous research on MT SAR sequences suggests that they increase the classification accuracy for all applications over single date images. Yet the presence of speckle noise remains a problem and all images in the sequence must be speckle filtered before acceptable classification accuracy can be attained. Several speckle filters designed specifically for MT sequences have been reported in the literature. Filtering in the spatial domain, as is usually done, reduces the effective spatial resolution of the filtered image. MT speckle filters operate in both the spatial and temporal dimensions, thus the reduction in resolution is not likely to be as severe (although a comparison between MT and spatial filters has not been reported). While this advantage may be useful when extracting spatial features from the image sequence, it is not quite as apparent for classification applications. This research explores the relative performance of spatial and MT speckle filtering for a particular classification application: mapping boreal forest types. We report filter performance using the radiometric resolution as measured by the equivalent number of looks (NL), and classification performance as measured by the classification accuracy. We chose representative spatial and MT filters and found that spatial speckle filters offer the advantage of higher radiometric resolution and higher classification accuracy with lower algorithm complexity. Thus, we confirm that MT filtering offers no advantage for classification applications; spatial speckle filters yield higher overall performance.     

Fast multisensor infrared image super-resolution scheme with multiple regression models

High resolution (HR) infrared (IR) images play an important role in many areas. However, it is difficult to obtain images at a desired resolution level because of the limitation of hardware and image environment. Therefore, improving the spatial resolution of infrared images has become more and more urgent. Methods based on sparse coding have been successfully used in single-image super-resolution (SR) reconstruction. However, the existing sparse representation-based SR method for infrared (IR) images usually encounter three problems. First, IR images always lack detailed information, which leads to unsatisfying IR image reconstruction results with conventional method. Second, the existing dictionary learning methods in SR aim at learning a universal and over-complete dictionary to represent various image structures. A large number of different structural patterns exist in an image, whereas one dictionary is not capable of capturing all of the different structures. Finally, the optimization for dictionary learning and image reconstruction requires a highly intensive computation, which restricts the practical application in real-time systems. To overcome these problems, we propose a fast IR image SR scheme. Firstly, we integrate the information from visible (VI) images and IR images to improve the resolution of IR images because images acquired by different sensors provide complementary information for the same scene. Second, we divide the training patches into several clusters, then the multiple dictionaries are learned for each cluster in order to provide each patch with a more accurate dictionary. Finally, we propose an method of Soft-assignment based Multiple Regression (SMR). SMR reconstructs the high resolution (HR) patch by the dictionaries corresponding to its K nearest training patch clusters. The method has a low level of computational complexity and may be readily suitable for real-time processing applications. Numerous experiments validate that this scheme brings better results in terms of quantization and visual perception than many state-of-the-art methods, while at the same time maintains a relatively low level of time complexity. Since the main computation of this scheme is matrix multiplication, it will be easily implemented in FPGA system.     

基于深度卷积神经网络的航空器检测与识别

针对军用机场大尺寸卫星图像中航空器检测识别的具体应用场景,建立了一套实时目标检测识别框架,将深度卷积神经网络应用到大尺寸图像中的航空器目标检测与识别任务中。首先,将目标检测的任务看成空间上独立的bounding-box的回归问题,用一个24层卷积神经网络模型来完成bounding-box的预测;然后,利用图像分类网络来完成目标切片的分类任务。大尺寸图像上的传统目标检测识别算法通常在时间效率上很难突破,而基于卷积神经网络的航空器目标检测识别算法充分利用了计算硬件的优势,大大缩短了任务耗时。在符合应用场景的自采数据集上进行测试,所提算法目标检测实时性达到平均每张5.765 s,在召回率65.1%的工作点上达到了79.2%的精确率,分类网络的实时性达到平均每张0.972 s,Top-1错误率为13%。所提框架在军用机场大尺寸卫星图像中航空器检测识别的具体应用问题上提出了新的解决思路,同时保证了实时性和算法精度。     

TM and IRS-1C-PAN data fusion using multiresolution decomposition methods based on the 'a tròus' algorithm

Image fusion represents an important tool for remote sensing data elaborations. This technique is used for many purposes. Very often it is used to produce improved spatial resolution. The most common situation is represented by a pair of images: the first acquired by a multispectral sensor with a pixel size greater than the pixel size of the second image given by a panchromatic sensor (PAN). Starting from these images fusion produces a new multispectral image with a spatial resolution equal, or close, to that of the PAN. Very often fusion introduces important distortions on the pixel spectra. This fact could compromise the extraction of information from the image, especially when using an automatic algorithm based on spectral signature such as in the case of image classification. In this work we present the analysis of two fusion methods based on multiresolution decomposition obtained using the 'a tròus' algorithm and applied to a pair of images acquired by Thematic Mapper (TM) and Indian Remote Sensing (IRS)-1C-PAN sensors. The methods studied are also compared with two classical fusion methods, the intensity, hue and saturation (IHS) and standardized principal components (SPC). Fused results are studied and compared using various tests including supervised classification. Most of the tests used have been extracted from literature regarding the assessment of spatial and spectral quality of fused images. This study shows that the methods based on multiresolution decomposition outperform the classical fusion methods considered with respect to spectral content preservation. Moreover, it is shown that some of the quality tests are more significant than others. The discussion of this last aspect furnishes important indications for data quality assessment methods.     

近红外高光谱图像数据预测技术

目的 受到传感器光谱响应范围的影响,可见光区域和近红外区域（400~2 500 nm）的高光谱数据通常使用不同的感光芯片进行成像,现有这一光谱区域典型的高光谱成像系统,如AVIRIS （airborne visible infrared imaging spectrometer）成像光谱仪,通常由多组感光芯片组成,整个成像系统成本和体积通常比较大,严重限制了该谱段高光谱探测技术的发展。为了能够扩展单感光芯片成像系统获得的高光谱图像的光谱范围,本文探索基于卷积神经网络的近红外光谱数据预测技术。方法 结合AVIRIS成像光谱仪的光谱配置,设计了基于残差学习的红外谱段图像预测网络,利用计算成像的方式从可见光范围的高光谱图像预测出近红外波段的光谱图像,并在典型的卫星高光谱遥感数据上进行红外光谱预测重构和基于重构的数据分类实验,以验证论文提出的红外光谱数据预测技术的可行性以及有效性。结果 本文设计的预测网络在Cuprite数据集上得到的预测近红外图像峰值信噪比为40.145 dB,结构相似度为0.996,光谱角为0.777 rad;在Salinas数据集上得到的预测近红外图像峰值信噪比为39.55 dB,结构相似性为0.997,光谱角为1.78 rad。在分类实验中,相比于只使用可见光图像,利用预测的近红外图像使得支持向量机（support vector machine,SVM）的准确率提升了0.6%,LeNet的准确率提升了1.1%。结论 基于AVIRIS传感器获取的两组典型卫星高光谱数据实验表明,本文提出的红外光谱数据预测技术不仅可基于计算成像的方式扩展可见光光谱成像系统的光谱成像范围,对于减小成像系统体积和质量具有重要意义,而且可有效提高可见光区域光谱图像数据在典型应用中的处理性能,对于提高高光谱数据处理精度提供新的技术支撑。     

Assessing the quality of temporal high-resolution classifications with low-resolution satellite image time series

Upcoming temporally and spatially high-resolution satellites such as Venus, SENTINEL-2, and Landsat Data Continuity Mission (LDCM) will provide very valuable data for land-cover and vegetation monitoring. However, owing to cloud cover and even to some rapid changes, a higher temporal resolution may be needed for some applications. In this work, we propose using the higher temporal resolution of satellites with mid to low spatial resolutions such as the upcoming PROBA-V. The aim of this work is to study how images provided by satellites with a lower spatial resolution but with a higher temporal one can be used to obtain information about the temporal classification derived from satellite image time series (SITS) provided by high-resolution satellites such as Venus, SENTINEL-2, or LDCM.

We show that the low-spatial-resolution SITS can be used to inform about the stability and relevance of high-spatial-resolution classification. Experiments include a wide variety of resolution ratios and study the use of each ratio from global to class-specific invalidation of the high-resolution classification map (computed from the high-spatial-resolution SITS).

This work contributes to the assessment of the usefulness of the joint use of PROBA-V data and Venus/SENTINEL-2/LDCM images for land-cover monitoring.     

Range map superresolution-inpainting,and reconstruction from sparse data

Range images often suffer from issues such as low resolution (LR) (for low-cost scanners) and presence of missing regions due to poor reflectivity, and occlusions. Another common problem (with high quality scanners) is that of long acquisition times. In this work, we propose two approaches to counter these shortcomings. Our first proposal which addresses the issues of low resolution as well as missing regions, is an integrated super-resolution (SR) and inpainting approach. We use multiple relatively-shifted LR range images, where the motion between the LR images serves as a cue for super-resolution. Our imaging model also accounts for missing regions to enable inpainting. Our framework models the high resolution (HR) range as a Markov random field (MRF), and uses inhomogeneous MRF priors to constrain the solution differently for inpainting and super-resolution. Our super-resolved and inpainted outputs show significant improvements over their LR/interpolated counterparts. Our second proposal addresses the issue of long acquisition times by facilitating reconstruction of range data from very sparse measurements. Our technique exploits a cue from segmentation of an optical image of the same scene, which constrains pixels in the same color segment to have similar range values. Our approach is able to reconstruct range images with as little as 10% data. We also study the performance of both the proposed approaches in a noisy scenario as well as in the presence of alignment errors.     

机载多传感器目标识别算法研究

主要目的是以机载探测设备为平台，针对机载探测设备自身特性，来设计一种更为有效的融合算法，来对敌机中的危险目标进行识别，在主要方法上运用神经网络技术、Dempster--Shafer（D—S）证据理论将来自于机载SAR雷达、机载前视红外搜索跟踪系统（IRS），电子支援措施（ESM）等探测设备多次观察所得到的数据，进行实时的时域和空域融合，对于来自于地面的电子情报（ELINT）的信息使用主观贝叶斯方法来同机载系统融合后的信息进行融合，从而达到准确的目标识别；最后通过实例仿真证明该算法适合于不同类型传感器不同格式信息之间的融合，其不仪能够适合于复杂的信号环境，并且在观测噪声比较大时，具有优良的性能和广泛的适应性。     

两级上下文卷积网络宽视场图像小目标检测方法

目标检测和识别已经在输电线路巡检中被广泛采用。由于图像数据量大,小目标分辨率低,现有的图像金字塔、特征金字塔和多异构特征融合等方法虽能准确地检测目标,却非常耗时,因而快速、准确地检测宽视场图像中小目标仍是一个挑战。此算法提出一个两个Faster-RCNs级联的上下文宽视场小目标检测卷积网络,首先,针对降分辨率的宽视场图像,利用一个Faster R-CNN来检测目标的上下文区域,然后,针对上下文区域对应的高分辨率原始图像,利用Faster R-CNN来检测来小目标。我们用航拍输电线路图像数据集进行了目标检测试验,试验结果表明,小目标检测方法达到了88%的检测精度,比单级Faster R-CNN检测方法具有更高的准确率。     

Single tree species classification with a hypothetical multi-spectral satellite

The requirements for high resolution multi-spectral satellite images to be used in single tree species classification for forest inventories are investigated, especially with respect to spatial resolution, sensor noise and geo-registration. In the hypothetical setup, a 3D tree crown map is first obtained from very high resolution panchromatic aerial imagery and subsequently each crown is classified into one of a set of known tree species such that the difference between a model multi-spectral image generated from the 3D crown map and an acquired multi-spectral satellite image of the forested area is minimized. The investigation is conducted partly by generating synthetic data from a 3D crown map from a real mixed forest stand and partly on hypothetical high resolution multi-spectral satellite images obtained from very high resolution colour infrared aerial photographs, allowing different hypothetical spatial resolutions. Conclusions are that until a new generation of even higher resolution satellites becomes available, the most feasible source of remote sensing data for single tree classification will be aerial platforms.     

Petascale large eddy simulation of jet engine noise based on the truncated SPIKE algorithm

With the emergence of petascale computing platforms, high-fidelity computational aeroacoustics (CAA) simulation has become a feasible, robust and accurate tool that complements theoretical and empirical approaches in the prediction of sound levels generated by aircraft airframes and engines. Differentiating itself from the broader discipline of computational fluid dynamics, CAA is particularly challenging as it demands high accuracy, good spectral resolution, and low dispersion and diffusion errors from the underlying numerical methods. Large eddy simulation based on space-implicit high-order compact finite difference schemes has been shown to meet such stringent requirements. In this paper, we discuss a new, scalable parallelization scheme with a three-dimensional computational space partitioning. Unlike many traditional multiblock computational fluid dynamics (CFD) methods, our partitioning is non-overlapping. We use the truncated SPIKE algorithm to solve the governing equations accurately and limit one-sided biased differentiation to just the physical boundaries. We present experimental performance data collected on Kraken and Ranger, two near-petascale computing platforms.     

一种红外成像检测系统设计

红外成像系统的性能检测是产品质量的重要保证,本文介绍了一种红外成像系统的检测设备的设计,通过设计一定特征的红外目标图像,检测红外成像系统的成像功能,同时检测红外成像系统的分辨率、成像坏元等关键性指标,通过检测系统控制设计保证检测系统输出的红外图像的温度阈值范围和质量满足被测对象的要求,并能够提供可设置的图像,实现了对红外成像系统的功能和性能检测。检测系统温度分辨率不大于0.2度,图像相对畸变2%,目标运动精度不大于0.05度,光能分布不均匀度不大于10%,视场角不小于8度。     

环境一号B星红外多光谱相机综合辐射定标

环境一号B星红外多光谱相机(简称HJ-1B星IRS)在轨运行过程中受各种因素的影响,其探测性能不断发生衰减,绝对辐射定标可有效地改善遥感器辐射特性的测量精度,从而提高遥感数据辐射质量.文中针对HJ-1B星IRS星上定标系统特点,利用2010年8月地面同步试验数据对IRS热红外通道进行综合辐射定标,修正了原始定标系数,并进行定标系数的真实性检验.结果表明借助地面试验数据进行的综合辐射定标弥补了星上定标系统的不足,获得的定标系数准确可靠,可满足遥感数据定量化应用的需要.