机器学习在岩矿地球化学研究中的应用——综述与思考

岩石与矿物的地球化学成分数据具有高维度特征。传统的岩矿地球化学成分研究主要采用二元/三元图解判别法,准确率不高,在数理统计方法上有欠缺。机器学习方法非常适用于对大样本高维度的岩矿成分数据进行数理统计处理。笔者等在介绍机器学习常见算法基本原理的基础上,总结近5年来国内外学者将机器学习方法应用于岩石矿物成分数据研究的实例,包括：(1)根据矿物成分溯源其母岩(源岩)、判别矿床类型,(2)新生代火山岩溯源,(3)判别变质岩原岩,(4)依据岩浆岩成分判别大地构造环境等。已有的研究实例显示,机器学习方法的准确度明显优于传统的低维度判别法。机器学习本质是分析大样本数据的高维度变量之间的相关、归类等多元统计问题。推广机器学习的应用需要建设开放获取(Open Access)的矿物、岩石成分数据库,同时全面实施开放研究(Open Research)的发表策略。     

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。     

Uncertainty pattern in landslide susceptibility prediction modelling: Effects of different landslide boundaries and spatial shape expressions

In some studies on landslide susceptibility mapping (LSM), landslide boundary and spatial shape characteristics have been expressed in the form of points or circles in the landslide inventory instead of the accurate polygon form. Different expressions of landslide boundaries and spatial shapes may lead to substantial differences in the distribution of predicted landslide susceptibility indexes (LSIs); moreover, the presence of irregular landslide boundaries and spatial shapes introduces uncertainties into the LSM. To address this issue by accurately drawing polygonal boundaries based on LSM, the uncertainty patterns of LSM modelling under two different landslide boundaries and spatial shapes, such as landslide points and circles, are compared. Within the research area of Ruijin City in China, a total of 370 landslides with accurate boundary information are obtained, and 10 environmental factors, such as slope and lithology, are selected. Then, correlation analyses between the landslide boundary shapes and selected environmental factors are performed via the frequency ratio (FR) method. Next, a support vector machine (SVM) and random forest (RF) based on landslide points, circles and accurate landslide polygons are constructed as point-, circle- and polygon-based SVM and RF models, respectively, to address LSM. Finally, the prediction capabilities of the above models are compared by computing their statistical accuracy using receiver operating characteristic analysis, and the uncertainties of the predicted LSIs under the above models are discussed. The results show that using polygonal surfaces with a higher reliability and accuracy to express the landslide boundary and spatial shape can provide a markedly improved LSM accuracy, compared to those based on the points and circles. Moreover, a higher degree of uncertainty of LSM modelling is present in the expression of points because there are too few grid units acting as model input variables. Additionally, the expression of the landslide boundary as circles introduces errors in measurement and is not as accurate as the polygonal boundary in most LSM modelling cases. In addition, the results under different conditions show that the polygon-based models have a higher LSM accuracy, with lower mean values and larger standard deviations compared with the point- and circle-based models. Finally, the overall LSM accuracy of the RF is superior to that of the SVM, and similar patterns of landslide boundary and spatial shape affecting the LSM modelling are reflected in the SVM and RF models.     

基于全卷积网络的高分辨遥感影像目标检测

目标检测是遥感图像分析处理中的研究热点之一,具有十分重要的科研和应用价值。传统遥感影像目标检测方法多使用人工构造的浅层次特征,结合支持向量机、随机森林、Adaboost等分类器进行目标识别,难以充分挖掘和利用影像中的深层特征。近年来,深度学习,特别是卷积神经网络在图像认知方面取得了巨大成功。在目标检测领域,以Faster R-CNN算法为代表的方法取得了突破性进展,检测精度大幅提高,检测速度达到了近实时的性能。但是,Faster R-CNN算法由于使用了感兴趣区域（RoI）池化层,各个RoI计算不共享,因此检测速度依然有待提高。R-FCN基于全卷积网络结构,同时采用位置敏感池化来引入平移变化,抵消全卷积网络造成的平移不变形问题,检测精度和效率都有了很大的提高。本文阐述了R-FCN算法原理,并运用于高分辨遥感影像目标检测分析了不同参数和网络结构对R-FCN检测效果的影响,比较了利用Fast R-CNN、Faster R-CNN和R-FCN 3种算法进行飞机识别的性能。试验结果表明,利用R-FCN进行飞机识别定位可以达到99.3%的准确率和每张图180 ms的检测速度。     

Performance analysis of different predictive models for crop classification across an aridic to ustic area of Indian states

The purpose of this study is to present comparative performance analysis of different machine learning algorithms for large area crop classification. Ten Indian districts with significant rabi crops viz. wheat, mustard, gram, red lentils (masoor) have been selected for the study. Most popular classical ensemble models – bagging/ARCing, random forest (RF), gradient boosting and Importance Sampled Learning Ensemble (ISLE) with traditional single model (decision tree) have been selected for comparative analysis. To incorporate dependency of large area crop in different variables viz. parent material and soil, phenology, texture, topography, soil moisture, vegetation, climate etc., 35 digital layers are prepared using different satellite data (ALOS DEM, Landsat-8, MODIS NDVI, RISAT-1, Sentinental-1A) and climatic data (precipitation, temperature). In rabi season, field survey about crop type is carried out to prepare training data. Performance is evaluated on the basis of marginal rates, F-measure and Jaccard’s coefficient of community, Classification Success Index and Agreement Coefficients. Score is calculated to rank the algorithm. RF is best performer followed by gradient boosting for crop classification. Other ensemble methods ARCing, bagging and ISLE are in decreasing order of performance. Traditional non-ensemble method decision tree scored higher than ISLE.     

Learning to combine multiple string similarity metrics for effective toponym matching

Several tasks related to geographical information retrieval and to the geographical information sciences involve toponym matching, that is, the problem of matching place names that share a common referent. In this article, we present the results of a wide-ranging evaluation on the performance of different string similarity metrics over the toponym matching task. We also report on experiments involving the usage of supervised machine learning for combining multiple similarity metrics, which has the natural advantage of avoiding the manual tuning of similarity thresholds. Experiments with a very large dataset show that the performance differences for the individual similarity metrics are relatively small, and that carefully tuning the similarity threshold is important for achieving good results. The methods based on supervised machine learning, particularly when considering ensembles of decision trees, can achieve good results on this task, significantly outperforming the individual similarity metrics.     

一种优化的Faster R-CNN小目标检测方法

图像目标检测是计算机视觉与数字图像处理的一个热门方向,其主要任务是找出图像中感兴趣的物体并确定物体的位置与类别。目前基于深度学习模型是主流的目标检测算法,利用其解决诸多学科问题成为一种趋势。本文采用区域卷积神经网络（Faster R-CNN）深度学习算法和相关图像处理算法,以ResNet50、ResNet101为骨干网络,采用特征金字塔网络开展新冠疫情期间武汉市车辆变化监测,以此分析疫情下的武汉市内部活动强度。结果显示：本文车辆目标检测方法的精确率为0.96,召回率为0.92,平均精度为0.85。疫情前（2019年11月17日）、中（2020年02月22日）车辆变化情况为：武汉汇聚中心分别为263、32辆,汪家嘴立交桥分别为89、44辆,新兴工业园分别为554、347辆,经开未来城分别为188、57辆。可知,疫情导致武汉市人口出行减少、车辆活动明显降低。     

多特征流形鉴别嵌入的高分辨率遥感影像分类

针对高分辨遥感影像同谱异物、同物异谱导致单一特征分类结果精度较差的问题,本文提出了多特征流形鉴别嵌入的高分辨率遥感影像分类方法。该方法首先提取高分辨率影像数据的光谱特征与LBP纹理特征;然后通过样本数据的联合光谱、纹理特征的空间距离及对应的类别信息,构建影像对象的类间图与类内图,用于学习高分辨率影像上的鉴别流形结构,保证在嵌入空间上尽可能不同地物特征分离、相同地物特征紧聚,确保相同地物光谱、纹理特征的相似性,完成光谱、纹理鉴别特征的有效提取,以充分挖掘影像特征,有效提高影像的分类精度。在GF-2遥感数据集上进行试验,结果表明本文算法可实现多特征的有效融合,分类精度均优于传统方法,可达93.41%。     

利用深度学习模型智能识别地下排水管道缺陷

排水管道健康状况直接影响整个城市的排水效果,CCTV检测作为目前最为常见的排水管道健康状况检测方法,仍存在自动化程度不高、工作效率低下、严重依赖人工经验等问题。为解决以上问题,本文将当前先进的深度学习技术与地下排水管道缺陷检测技术相结合,提出了一种基于深度学习模型的地下排水管道缺陷智能识别技术。同时,将管道缺陷智能识别与管道检测工作流程紧密结合,实现城市排水管道检测报告的快速自动生成等功能,从而大大提高排水管道缺陷的检测效率。