江苏沿海滩涂利用模式遥感影像解译和时空变化分析

为了弄清江苏沿海滩涂利用模式分布情况,定量分析其时空变化趋势,本文基于Google Earth Engine云计算平台和Landsat卫星遥感影像数据,结合随机森林分类方法,通过机器学习建立了滩涂利用模式自动识别分类方法,样本训练后识别准确率达93％以上.利用该方法提取了2002年、2014年和2019年江苏沿海的土地...     

一种基于改进CoHOG的视觉SLAM算法

移动机器人在SLAM的闭环检测环节计算量大、运行时间长、匹配误差大,从而导致闭环检测精度较低。针对该问题,本文在CoHOG闭环检测算法的基础上进行改进,将算法中的HOG描述符改进为GDF-HOG描述符,以增强图像特征表现,提高图像特征提取效率;在匹配环节前添加GDF-HOG全局粗匹配,以减少视觉模板的数量,提高算法的计算效率;在匹配环节后添加感兴趣区域（ROI）位置匹配进行检验,以减少闭环检测的假阳性,提高准确率。将本文闭环检测算法与RatSLAM相结合,在公开数据集与真实环境中进行测试,测试结果表明,本文算法在闭环检测环节的准确率较高,且对环境的适应能力较强。     

塔什干低涡天气过程的计算机识别

根据塔什干低涡的定义,利用NCEP／NCAR日平均再分析高度场、温度场资料,运用计算机编程自动识别、追踪塔什干低涡的中心位置、中心强度和持续时间,进而获取塔什干低涡天气过程的数据集。此计算机识别程序中有两个关键判断条件：一是某格点i的高度值小于其外围格点高度值,二是低涡范围内任意格点的纬向温度二阶导数大于零。实践证明该方法对塔什干低涡系统的识别准确且省时省力,同时避免了人工挑选的主观性误差。     

颜色和形态学在交通标志识别中的应用研究

从自然场景中识别出交通标志是智能交通系统中的一个重要问题,本文提出了一种基于颜色和形态学的交通标志识别方法.该方法在HSV彩色空间进行颜色阈值分割,并采用形态学重构提取目标区域.再通过标记法检测目标区域形状,检测出交通标志.最后,利用SIFT算法进行模板匹配,识别出交通标志.实验证明,该方法在识别正确率和运行效率上具有明显优势.     

地名识别与匹配的概率统计方法

建立了一个试验用地名库和地理语料库,在此基础上构建对地名用字可信度的统计分析模型。通过分析地名在中文文档中的使用习惯和规律,总结出经常与地名一起使用的且具有地名指示含义的辅助字或词,以此为基础建立地名识别辅助词词库和地名识别的规则库。对地名库和地理语料库的用字进行统计分析,通过设定地名用字可信度概率阈值和辅助词指示作用对文本中潜在地名进行初步的筛选形成候选地名;在粗筛选产生的候选地名基础上结合地名识别规则进一步确认,以提高地名识别的准确率。     

Developing the tools for geological shape analysis,with regional‐ to local‐scale examples from the Kalgoorlie Terrane of Western Australia

Geological map data are often underused in mineral‐exploration programs, which rely increasingly on regolith geochemistry and geophysical and other remotely sensed data to generate exploration targets. However, solid geology maps, which are progressively being upgraded due to improved interpretations of superior, remotely sensed images and airborne geophysical data, can be useful in targeting specific types of mineral deposits, which formed late in the evolutionary history of the host terrane. In such terranes, the present map geometry is essentially the same as that at the time of deposit formation. This is the case for orogenic lode‐gold deposits, which commonly show predictable structural controls and/or structural geometry. Thus, the shape of a rock body, or combinations of structures and rock bodies, may provide an important guide to the exploration potential for orogenic lode‐gold deposits. However, until recently, there has been a dearth of techniques to quantify the various properties of shape, and hence test the potential of the two‐dimensional shape of geological bodies in map view as an exploration tool. Integrating techniques from the field of pattern recognition with a modern Geographical Information System (GIS) can provide the shape‐analysis tools required to investigate the geometries of geological shapes. Two‐dimensional shape analysis is now possible through the calculation of several shape metrics including, but not restricted to, aspect ratio, blockiness, elongation, compactness, complexity, roundness, spreadness and squareness. Methods are developed for describing the geometries of rock units about mineral deposits, or any geological features, at any scale, which for the first time makes it possible to compare shapes. These shape‐analysis techniques are tested using orogenic lode‐gold deposits, particularly those in the Kalgoorlie Terrane of the highly auriferous Late Archaean Norseman‐Wiluna Belt of Western Australia. On a global scale, shape analysis indicates that those greenstone belts whose volcanic rock sequences have high elongation and relative low roundness, complexity and aspect ratio (e.g. Kalgoorlie Terrane) are likely to be the most richly endowed in gold. On a more local scale, characteristics of the shape of geological features around the Golden Mile deposit are calculated and used to test the likelihood of occurrence of gold deposits with similar geometry elsewhere in the Kalgoorlie Terrane. The area with the most closely matching shape, on the basis of a 2 km clipping‐circle radius, chosen on the basis of available proximity‐analysis data, corresponds to the recently discovered Ghost Crab deposit, illustrating the potential of the shape analysis methodology in mineral exploration. Shape analysis is, at least in part, scale dependent, due to the inherent problem of being able to define rock boundaries more precisely in units that have strong geophysical signatures than those with weak signatures in poorly exposed terranes. Overcoming this problem is a challenge to the application of this methodology.     

Comprehensive Assessment of Seawater Quality on an Improved Attribute Recognition Model

The attribute recognition model （ARM） has been widely used to make comprehensive assessment in many engineering fields, such as environment, ecology, and economy. However, large numbers of experiments indicate that the value of weight vector has no relativity to its initial value but depends on the data of Quality Standard and actual samples. In the present study, the ARM is enhanced with the technique of data driving, which means some more groups of data from the Quality Standard are selected with the uniform random method to make the calculation of weight values more rational and more scientific. This improved attribute recognition model （IARM） is applied to a real case of assessment on seawater quality. The given example shows that the IARM has the merits of being simple in principle, easy to operate, and capable of producing objective results, and is therefore of use in evaluation problems in marine environment science.     

用雷达回波资料识别雹云的最新进展

本文评述了当前国内外利用雷达回波资料识别雹云的现状,着重介绍了笔者近年来利用雷达回波资料建立雹云识别模式的研究工作,并展望了雹云识别的未来发展前景。     

EXPLORING SPATIAL FAMILIARITY*

Spatial familiarity is a natural language concept that has no well-defined technical interpretation. We explore here some possible dimensions of the concept, provide subjective measurements of the dimensions, and evaluate their independence. The four dimensions examined are locational knowledge, visual recognition, name identification, and interaction frequency. We suggest that both cognitive and behavioral dimensions are embedded in the concept of spatial familiarity and must be part of the meaning of the concept and of any attempt to make it operational.     

来自大别山深成侵入岩图像深度迁移学习的可解释性研究

岩石图像识别是以深度学习为代表的感知智能在地质领域的典型应用场景。已有研究显示网络结构简单的深度卷积神经网络能够在岩石图像上取得比复杂网络结构高的分类准确率。这与ImageNet数据集上网络结构越深越好的趋势相悖。如何解释这一现象？深成侵入岩为显晶质,自形—半自形粒状结构,块状构造,其分类的依据是其矿物成分及相对含量。大别山地区岩浆活动广泛,中生代深成侵入岩广泛出露。岩石类型包括超镁铁质岩类、辉长岩类、闪长岩类、正长岩类、二长岩类和花岗岩类,基本覆盖IUGS推荐的深成侵入岩分类方案中的岩石类型。选取大别山地区中生代深成岩图像开展不同网络结构预训练模型迁移学习对比试验,能够专注于深度学习对矿物成分特征的学习解释,降低构造因素的影响。借助局部可理解的模型解释技术和特征图可视化技术,分别从全连接层分类决策区域可视化和卷积隐层可视化两方面对深度学习模型开展可解释性研究。结果表明简单网络结构的卷积神经网络能够提取不同矿物所表现出的颜色特征以及不同矿物组合所表现出的纹理特征。AlexNet模型的削减试验进一步证明：对于岩石图像深度学习,网络结构并不总是越深越好。