图像自动识别技术在海洋浮游生物分析中的应用

介绍了一种利用数学形态学特征和Gabor纹理特征,结合主成分分析与支持向量机对胶州湾沿岸7种浮游生物的活体图像进行自动识别的方法。实验结果表明,基于主成分分析的降维识别模式可以提高系统识别性能,其平均识别正确率达78.5%,通过对图像采集、图像处理、特征的选取等方面做进一步的改进和提高,基于计算机数字图像的海洋浮游生物自动识别方法将为海洋生态环境监测提供新的实时、快速、高效检测平台。     

Computer vision techniques for quantifying, tracking, andidentifying bioluminescent plankton

This paper applies computer vision techniques to underwater video images of bioluminescent biota for quantifying, tracking, and identification. Active contour models are adapted for computerized image segmentation, labeling, tracking, and mapping of the bioluminescent plankton recorded by low-light-level video techniques. The system automatically identifies luminous events and extracts features such as duration, size, and coordinates of the point of impact, and uses this information to taxonomically classify the plankton species. This automatic classification can aid oceanographic researchers in characterizing the in situ spatial and temporal relationships of these organisms in their underwater environment. Experiments with real oceanographic data are reported. The results indicate that the approach yields performance comparable to human expert level capability. Furthermore, because the described technique has the potential to rapidly process vast quantities of video data, it may prove valuable for other similar applications     

多维信息在改善沿海地区图像分类精度中的应用方法

在某些专题研究中,由于研究目标、大气传输、下垫面及其他因素的限制,波谱分析的统计分类模式并不能完全满足研究需要。虽然多时相高精度图像可以提高分类精度,但常因其费用相当昂贵而难以实现。因此,如何在现有资料基础上,提取更多有效信息是亟需解决的问题。针对沿海地区特有的复杂地理现象,以TM图像为基础,通过量化手段进行最佳信息维数的确定与最佳信息维数组合的选择,探讨了信息维数对分类精度的影响。结果表明,在原有图像基础上,进行信息挖掘并重新组合可以得到能满足工作需要的分类精度,组合维数以6维最佳,更多的信息融合对精度影响不明显。     

Toward remote seafloor classification using the angular response ofacoustic backscattering: a case study from multiple overlapping GLORIAdata

While the average seafloor backscatter strength within a narrow range of grazing angles can be used as a first-order classification tool, this technique often fails to distinguish seafloors of known differing geological character. In order to resolve such ambiguities, it is necessary to examine the variation in backscatter strength as a function of grazing angle. For this purpose, a series of multiply overlapping GLORIA sidescan sonar images (6.5 kHz) have been obtained in water depths ranging from 1000 to 2500 m. To constrain the placement of acoustic backscatter measurements and to measure the true impinging angle of the incident wave, the corresponding seafloor was simultaneously surveyed using the Seabeam multibeam system. As a result of the multiple overlap, the angular response of seafloor backscatter strength may be derived for regions much smaller than the swath width. By using the derived angular response of seafloor backscatter strength in regions for which sediment samples exist, an empirical seafloor classification scheme is proposed based on the shape, variance, and magnitude of the angular response. Because of the observed variability in the shape of the angular response with differing seafloor types, routine normalization of single-pass swath data to an equivalent single grazing angle image cannot be achieved. As a result, for the case of single-pass surveys, confident seafloor classification may only be possible for regions approaching the scale of the swath width     

用于海底底质分类的多波束声强数据选取研究

高质量的海底声强图是进行多波束海底底质分类、目标识别的基础。要得到"单纯"反映海底底质信息的声强图,就需要对原始声强数据进行地形改正,消除地形因素的影响。在描述了多波束数据中水深数据不能满足声强数据的改正要求问题的基础上,提出了以水深数据覆盖范围为约束的声强数据选取方法。实例计算结果表明:该方法在能有效地选取高质量的声强数据,提高了基于声强图像的海底底质分类精度。     

基于数据挖掘的GF-1 遥感影像绿潮自适应阈值分区智能检测方法研究

由于受到云雾的影响,可见光影像能够高效用于绿潮检测的数据源较为有限,特别是云覆盖较为严重的可见光影像,基本无法用于检测绿潮。即使影像数据是在薄云、薄雾、无云覆盖的情况下获取的,由于其光谱反射值存在较大差异,依然很难采用同一阈值进行绿潮检测。基于此,为了提高可见光影像的利用率,实现不同云覆盖情况下,绿潮的高精度自适应阈值的自动检测,本文以GF-1影像为数据源,首先采用K-means聚类和C4.5决策树方法实现影像云覆盖情况的自动识别;其次,选取大量不同云覆盖情况下子图像样本（每个子图像样本中均包含绿潮和海水两类）,分析得出不同云覆盖情况下绿潮和海水的区分阈值y与影像光谱差x=band_nir-band_red之间所具有的线性关系;然后,利用分析得出的线性关系提出一种适用于GF-1影像的绿潮分区自适应阈值自动检测方法。最后,为验证提出方法的有效性,分别采用NDVI方法、EVI方法和本文提出的自适应阈值自动检测方法进行绿潮提取实验。实验结果表明,对于GF-1卫星遥感数据,本文提出的绿潮自适应阈值分区自动检测方法明显优于传统的NDVI和EVI检测方法,不仅提高了绿潮的监测精度,而且实现了绿潮提取的全自动化。     

图像融合在遥感中的应用

介绍了关于低分辨率的多光谱图像与高分辨率的全色图像相融合的一些方法。图像融合的目的是为了获取高分辨率的多光谱图像,它既包含了多光谱信息又具有全色图像的高空间分辨率特点。图像融合在遥感中的主要目的在于尽量保持图像的光谱信息．以便用于土地覆盖分类等领域,这一点与其在军事应用和计算机辅助设计领域有所不同。阐述了如何在不使用高分辨率的全色图像情况下进行融合的方法,由于同时获取的多光谱图像之间存在亚像元级的偏移现象,将位移量作为附加信息来有效提高采样频率,以此获取高空间分辨率信息。最后选取了SPOT卫星的多光谱图像作为例子．对文中所阐述的算法进行了试验,并对试验结果作了比较分析。     

应用卫星遥感边缘效应解译冬季黄海的水团配置

作者提出叶绿素是海洋浮游植物生物量的一个重指标,在两水系交汇的锋区,由于浮游植物大量繁殖,叶绿素富集,处于高生产力状态。同步获取黄海叶绿素在富集带及邻近外区的分布,可以确定水团边界。对此,作者根据卫星图象经几何纠正后,通过与黄海叶绿素a的边缘效应信息及区域分布态势比较并解译,将冬季黄海的水团划分为八种。本文还引证了遥感与其它常规资料佐证。     

珠江口夏季水体中的氮和磷

根据1999年7月17～28日于珠江口现场调查和实验的资料,研究夏季水体中氮、磷的分布、形态变化和初级生产力的限制因素.结果表明该海域氮含量高,N/P属于世界上高值区之一.从河口向外海运输过程中,氮和磷的形态和浓度均有剧烈的变化.虽然氮在中途中有新源的补充;但由于外海水的入侵稀释、生物吸收和形态变化的迁移作用,NO₃-和可溶无机氮的浓度总的变化趋势仍是随盐度增大而大幅度地降低,以至珠江口外出现N/P低于16.由于夏季水体层化稳定,在表、底层其生物地球化学变化方向相反,PO₄3的浓度变化互成镜像关系并可按盐度分为3段不同特征的反应区.初级生产力的限制因素在大部分区域是磷,但从口门至最大浑浊带和口外区则分别是浊度(或光照)及可溶无机氮.现场培养实验再现了真光层和底层氮和磷的生物地球化学过程差异并表明磷的循环和再生比氮迅速;在可溶无机氮浓度大且高N/P的海域,磷的再生可成为水华的引发因素,而氮被耗尽却是水华消亡的原因.总体上夏季该区水体氮的迁出率比磷高.于水体层化稳定的区域,氮和磷的生物地球化学作用在真光层以浮游生物吸收占优势、在下层以有机物的降解和可溶无机态的再生为主,当层化消失、上下水体充分混合则可完成循环.     

基于多尺度纹理和光谱信息的海岛土地利用遥感分类方法研究

本文基于灰度共生矩阵提取多尺度纹理指数影像,将纹理指数影像作为逻辑波段与原始影像波段组合,进行最大似然监督分类的实验,以改善单纯基于像元光谱影像分类技术的不足。利用SPOT-5卫星遥感影像对西门岛土地利用分类的实例进行分析。结果表明,由SPOT-5遥感影像第一波段和第二波段计算所得的相关性和同质性纹理指数影像以及由第三波段计算所得的均值纹理和第四波段计算的同质性纹理指数影像,对海岛土地利用分类精度的提高均有较为显著的效果,其中加入多尺度纹理信息后西门岛土地利用遥感影像分类的总体精度由75.41%提高到89.41%。     

基于DeepLabv3架构的高分辨率遥感图像分类

针对目前使用机器学习解决高分辨率遥感图像分类主要存在下采样导致的细节信息丢失问题,提出了一种基于DeepLabv3架构的小波域DeepLabv3-MRF(Markov random field,MRF)算法。选择当前较为普遍的DeepLabv3架构分类算法,能够获得更为精确的分类结果;采用小波域DeepLabv3-MRF算法,还能够获得更为清晰的边缘细节信息。选取南方某地区高分辨率无人机遥感图像进行分类实验,通过小波变换的方向性、非冗余性以及MRF变换像素空间的交互性这三个方面,将分类结果与原始DeepLabv3架构分类结果对比分析。结果表明,所提出的分类方法精度明显高于原始DeepLabv3架构分类算法的精度,总体精度可提升3%左右,并且可以充分表达高分辨率遥感图像细节信息。     

海洋观测图像的深度哈希检索及FPGA在轨实现

近年来,星载/机载的在轨海洋观测设备日益增多。如何从星/机在轨观测的数据中快速且准确地检索出关键图像并进行回传,逐渐成为在轨实时处理的重要研究课题。哈希函数作为一种映射函数,能够将不同图像变换为相同长度的二进制码,并且保持类似图像二进制码的相似性。为了保证二进制编码的特征表达能力,采用深度卷积神经网络(Convolutional Neural Network,CNN)提取海洋图像的特征信息,对图像进行哈希变换。本文以海洋观测图像作为训练数据集,基于深度卷积神经网络在GPU(Graphics Processing Unit)上训练哈希算法。在此基础上,本文将训练好的哈希函数在FPGA(Field Programmable Gate Array)上实现,完成海洋观测图像的在轨快速检索,为海洋观测实时处理提供了有效技术手段。     

基于双极化SAR图像的U-Net海冰多分类模型

北极多年冰在近几十年有明显的减少趋势,与北极海冰的厚度、体积和夏季最小海冰范围的减少密切相关。合成孔径雷达(synthetic aperture radar, SAR)具有全天时、全天候成像能力,基于SAR卫星影像的海冰分类对监测北极多年冰具有重要意义。基于深度学习U-Net模型,以SAR图像的双极化信息为模型输入,构建了像素级的海水、一年冰和多年冰多分类模型。与已有SAR图像海冰分类方法(支持向量机、随机森林和卷积神经网络)进行对比,基于双极化SAR图像的U-Net海冰分类模型的准确率、平均重叠度和Kappa系数,分别达到了90.73%、0.831和0.849,优于其他对比模型,分别提升了4.08%～19.04%, 0.063～0.321和0.111～0.335。此外,针对SAR图像水平-垂直极化(horizontal-vertical polarization, HV)有明显的条状热噪声和水平-水平极化(horizontal-horizontal polarization, HH)受入射角效应而亮度不均匀的特点,设计敏感性实验,研究HV噪声、入射角和灰度共生矩阵(gray leve...     

Use of SPOT 5 and IKONOS imagery for mapping biocenoses in a Tunisian Coastal Lagoon (Mediterranean Sea)

Mapping marine biocenoses is an efficient method for providing useful data for the management and conservation of Mediterranean lagoons. Fused images from two satellites, SPOT 5 and IKONOS, were tested as management tools for identifying specific ecosystems in the El Bibane lagoon, situated in southern Tunisia near the Libyan border. The objectives of this study were to provide a precise map of the entire El Bibane lagoon using fused images from SPOT 5 and to compare fused images from SPOT 5 and IKONOS over a test-area. After applying a supervised classification, pixels are automatically classified in four classes: low seagrass cover, high seagrass cover, superficial mobile sediments and deep mobile sediments. The maps of the lagoon revealed and confirmed an extremely wide distribution of seagrass meadows within the lagoon (essentially Cymodocea nodosa; 19 546 ha) and a large area of mobile sediments more or less parallel to the shore (3 697 ha). A direct comparison of overall accuracy between SPOT 5 over the entire area, SPOT 5 over the test-area and IKONOS over the test-area revealed that these tools provided accurate mapping of the lagoon environment (83.25%, 85.91% and 73.41% accuracy, respectively). The SPOT 5 images provided greater overall accuracy than the IKONOS image, but did not take into account the heterogeneous spatial structure of the seagrasses and sediments present in the lagoon environment. Although IKONOS imagery provided lower overall accuracy than SPOT 5, it proved a very useful tool for the mapping of heterogeneous structures as it enabled the patchiness of formations to be better taken into account. The use of SPOT 5 and IKONOS fused images appears to be very promising for completing the mapping of lagoons in other regions and countries of the Mediterranean Sea.     

协同主动学习和半监督方法的海冰图像分类

海冰遥感光谱影像分类中标签样本难以获取,导致海冰分类精度难以提高,但是大量包含丰富信息的未标签样本却没有得到充分利用,针对这种情况,提出一种协同主动学习和半监督学习方法用于海冰遥感图像分类。在主动学习部分,结合最优标号和次优标号、自组织映射神经网络以及增强的聚类多样性算法来选择兼具不确定性和差异性的样本参与训练;在半监督学习部分,利用直推式支持向量机,并且融合主动学习思想从大量未标签样本中选取相对可靠且包含一定信息量的样本进行迭代训练;然后协同主动学习分类结果和半监督分类结果,通过一致性验证保证所加入伪标签样本的正确性。为了验证方法的有效性,分别采用巴芬湾地区30 m分辨率的Hyperion高光谱数据(验证数据为15 m分辨率的Landsat-8数据)和辽东湾地区15 m分辨率的Landsat-8数据(验证数据为4.77 m分辨率的Google Earth数据)进行海冰分类实验。实验结果表明,相对其他传统方法,该协同分类方法可以在只有少量标签样本的情况下,充分利用大量未标签样本中包含的信息,实现快速收敛,并获得较高的分类精度(两个实验的总体精度分别为90.003%和93.288%),适...     

基于合成孔径雷达影像的海洋溢油纹理特征参数分析

溢油已是当前海洋生态环境破坏的主要因素之一,因此对海洋溢油的检测分析是当前海洋环境保护的一个重要课题。传统的溢油提取仅仅是单独依靠光学影像的光谱信息或者合成孔径雷达(SAR)影像的后向散射系数信息进行提取,这会造成很多同谱异物或者粗糙度相近似的地物错分,因此除了利用传统的影像信息以外,还需结合影像的纹理信息,从而提高溢油提取的精度,减少错分地物的数量。选用2006年渤海地区的三景同轨SAR影像作为数据基础,应用基于灰度共生矩阵的方法对其进行纹理分析。该方法可以很好地对图像区域和表面进行感知并能够从像元的灰度相关性上对纹理特征进行详细描述,因此适合于SAR影像的海洋溢油检测。在纹理分析的过程中有很多的参数需要选择,参数选择的好坏将直接影响最终提取结果的精度。通过对纹理分析过程中的参数进行讨论、实验、选择与验证,最终确定了基于灰度共生矩阵纹理分析中各参数的值,并选择了局部平稳、非相似性、对比度、变化量4个特征量作为溢油提取的纹理特征统计量。将纹理特征与SAR自身的后向散射系数相结合,通过神经网络分类法对其进行分类,并计算出分类精度为80.65%,分类效果良好。由此说明了将影像的传统信息与纹理信息相结合进行溢油提取是一种可行而有效的方法,同时也为后续的海洋溢油检测工作奠定了一定的基础。     

不同极化方式SAR与TM融合影像滨海湿地分类精度比较

以覆盖黄河口湿地区域的Radarsat-2 SAR全极化影像和Landsat-5 TM影像为例,将4种极化方式的SAR影像与TM影像分别进行融合,采用支持向量机对融合结果进行滨海湿地典型地物土地覆盖分类,并对分类结果进行比较评价,分析不同极化方式的SAR影像与TM影像融合结果在滨海湿地地区的分类能力。实验结果表明:采用与SAR影像融合的方法能够提高TM影像的分类精度,其中HV极化方式的SAR影像与TM影像的融合结果分类精度最高,最适用于滨海湿地土地覆盖分类研究。     

Rapid visualization of plankton abundance and taxonomic composition using the Video Plankton Recorder

Traditional methods for determining spatial distributions of planktonic taxa involve net, pump, and bottle collections followed by the tedious and time-consuming task of plankton sample analysis. Thus, plankton ecologists often require months or even years to process samples from a single study. In this paper, we present a method that allows rapid visualization of the distribution of planktonic taxa while at sea. Rapid characterization of plankton distributions is essential in the dynamic physical environment, where biological and physical patterns can change quickly. Such a “sample-and-observe” capability is necessary for mapping ephemeral features (such as patches, eddies, jets, plumes) and determining appropriate locations to conduct more localized sampling, including in situ observational studies. We describe the techniques used in imaging the plankton, analyzing the video, and visualizing the data. We present an example of at-sea data analysis conducted aboard R.V. Columbus Iselin on Georges Bank in May 1994 and visualizations of the 3-dimensional distribution of selected planktonic taxa in a 2 × 2 km × 90 m volume of seawater. A video of the image processing and visualization is included on the CD-ROM accompanying this volume and is an essential part of this paper.     

多波束声呐图像条带中央和边缘残差处理方法

多波束声呐图像是进行海底底质分类的主要数据源之一,由于受海洋噪声、声波散射和混响、仪器设备等因素影响,其经各项常规改正后仍存在明显残差,突出表现在中央波束区和条带重叠区,难以形成高质量的声呐图像。文中分析了多波束声呐图像残差的成因及影响,提出了一种基于多条带最小二乘拟合的多波束声呐图像残差处理方法。首先,得到相邻声脉冲(ping)信号中央区域、重叠区域以及整体趋势的拟合函数;然后,通过拟合函数计算得到中央和重叠区域的残差改正系数;最后,通过改正系数进行残差改正。实验分析表明,该方法在保留原始细节的基础上,有效削弱了残差对声呐图像的影响,对多波束声呐图像处理具有参考和应用价值。     

Analysis on coral reefs mapping using SPOT5 data at the Dongsha Atoll

Coral reefs are an sensitive-to-environment complex marine ecosystem. The ecosystem of corals is rich in biodiversity. Remote sensing offers a powerful tool for categorizing coral reefs and is the most cost-effective approach for the large-scale reef survey. The Dongsha Atoll, more than 300 km2 with an average depth of 10 m, is located at the northern continental margin of the South China Sea. It has been abused by destructive fishing during recent decades. Three satellite imageries (Quickbird2, ETM and SPOT5) are used to evaluate the capabilities of SPOT5 imagery to provide data that are useful for categorizing the current distribution of coral reefs therein. During the data processing, unsupervised classification functions are adopted for ETM and SPOT5 data, while the supervised classification method is used for Quickbird2. The classes are (or not) merged into coral reef, and then will be operated by vectorization, simplification, and topological analysis. There are 1 331 coral reefs larger than 100 m2 with a detection limit of 3×3 pixels at the multi-band data of Quickbird2, which is taken as the comparison baseline. The results extracted from SPOT5 and ETM images are less in number and area than those from the Quickbird2 image, whereas the results from SPOT5 data are better than those of ETM data at the silty lagoon due to its higher resolution. SPOT5 XS band 2 fails to distinguish the deep substrate inside the atoll compared with ETM data because of its poor penetration capability. Only SPOT XS band 1 cannot be used to differentiate coral reef from sand bottom. Merging the SPOT5 multi-bands data with the spatial resolution of SPOT5 pan-data and referring to ETM imagery are expected to provide an optimal satellite-based approach for mapping of coral reefs.