基于岩石图像深度学习的岩性自动识别与分类方法

岩石岩性的识别与分类对于地质分析极为重要,采用机器学习的方法建立识别模型进行自动分类是一条新的途径。基于Inception-v3深度卷积神经网络模型,建立了岩石图像集分析的深度学习迁移模型,运用迁移学习方法实现了岩石岩性的自动识别与分类。采用此方法对所采集的173张花岗岩图像、152张千枚岩图像和246张角砾岩图像进行了学习和识别分类研究,通过训练学习建立岩石图像深度学习迁移模型,并分别采用训练集和测试集中的岩石图像对模型进行了检验分析。对于训练集中的岩石图像,每组岩石分别用3张图像测试,三种岩石的岩性分类均正确,且分类概率值均达到90%以上,显示了模型良好的鲁棒性;对于测试集中的岩石图像,每组岩石分别采用9张图像进行识别分析,三种岩石的岩性分类均正确,并且千枚岩组图像分类概率均高于90%,但是花岗岩组2张图像和角砾岩组的1张图像分类概率值不足70%,概率值较其他岩石图像低,推测其原因是训练集中相同模式的岩石图像较少,导致模型的泛化能力减小。为了提高识别精确度,对准确率较低的岩石图像进行截取,分别取其中的3张图像加入训练集进行再训练,增加与测试图像具有相同模式的训练样本;在新的模型中,对3张图像进行二次检验,测试概率值均达到85%以上,说明在数据足够的状况下模型具有良好的学习能力。与传统的机器学习方法相比,所提出的岩石图像深度学习方法具有以下优点:第一,模型通过搜索图像像素点提取物体特征,不需要手动提取待分类物体特征;第二,对于图像像素大小,成像距离及光照要求低;第三,采用适当的训练集可获得较好的识别分类效果,并具有良好鲁棒性和泛化能力。     

卷积神经网络及其在矿床找矿预测中的应用——以安徽省兆吉口铅锌矿床为例

大数据人工智能地质学刚刚起步,基于大数据智能算法的地质研究是非常有意义的探索性实验。利用大数据和机器学习解决矿产预测问题,有助于人们克服不能全面考虑地质变量的困难及评估当前模型在已有数据中的可靠性。元素地表分布特征量主要受原岩成分、成矿作用影响和地表过程的影响,它们携带某些指示矿体就位的信息,即矿体在地下空间就位时在地表的响应,且未在地表过程中消失。以往的地球化学勘查工作仅仅识别异常,但未能发现矿体在地表响应的成矿特征量。本文以安徽省兆吉口铅锌矿床为例,通过机器学习,利用卷积神经网络算法,不断挖掘元素Pb分布特征与矿体地下就位空间的耦合相关性。经过1000次训练后,可以得到准确率0. 93,损失率0. 28的卷积神经网络模型。这种神经网络模型就是矿体在地下就位时元素在地表分布的响应,可以用来进行矿产资源预测。应用该模型对未知区进行预测,结果显示第53号区域具有很大概率存在尚未发现的矿体。     

Mapping fractional landscape soils and vegetation components from Hyperion satellite imagery using an unsupervised machine-learning workflow

An unsupervised machine-learning workflow is proposed for estimating fractional landscape soils and vegetation components from remotely sensed hyperspectral imagery. The workflow is applied to EO-1 Hyperion satellite imagery collected near Ibirací, Minas Gerais, Brazil. The proposed workflow includes subset feature selection, learning, and estimation algorithms. Network training with landscape feature class realizations provide a hypersurface from which to estimate mixtures of soil (e.g. 0.5 exceedance for pixels: 75% clay-rich Nitisols, 15% iron-rich Latosols, and 1% quartz-rich Arenosols) and vegetation (e.g. 0.5 exceedance for pixels: 4% Aspen-like trees, 7% Blackberry-like trees, 0% live grass, and 2% dead grass). The process correctly maps forests and iron-rich Latosols as being coincident with existing drainages, and correctly classifies the clay-rich Nitisols and grasses on the intervening hills. These classifications are independently corroborated visually (Google Earth) and quantitatively (random soil samples and crossplots of field spectra). Some mapping challenges are the underestimation of forest fractions and overestimation of soil fractions where steep valley shadows exist, and the under representation of classified grass in some dry areas of the Hyperion image. These preliminary results provide impetus for future hyperspectral studies involving airborne and satellite sensors with higher signal-to-noise and smaller footprints.     

Evaluation of effectiveness of three fuzzy systems and three texture extraction methods for building damage detection from post-event LiDAR data

Building damage maps after disasters can help us to better manage the rescue operations. Researchers have used Light Detection and Ranging (LiDAR) data for extracting the building damage maps. For producing building damage maps from LiDAR data in a rapid manner, it is necessary to understand the effectiveness of features and classifiers. However, there is no comprehensive study on the performance of features and classifiers in identifying damaged areas. In this study, the effectiveness of three texture extraction methods and three fuzzy systems for producing the building damage maps was investigated. In the proposed method, at first, a pre-processing stage was utilized to apply essential processes on post-event LiDAR data. Second, textural features were extracted from the pre-processed LiDAR data. Third, fuzzy inference systems were generated to make a relation between the extracted textural features of buildings and their damage extents. The proposed method was tested across three areas over the 2010 Haiti earthquake. Three building damage maps with overall accuracies of 75.0%, 78.1% and 61.4% were achieved. Based on outcomes, the fuzzy inference systems were stronger than random forest, bagging, boosting and support vector machine classifiers for detecting damaged buildings.     

Supervised classification of electric power transmission line nominal voltage from high-resolution aerial imagery

For many researchers, government agencies, and emergency responders, access to the geospatial data of US electric power infrastructure is invaluable for analysis, planning, and disaster recovery. Historically, however, access to high quality geospatial energy data has been limited to few agencies because of commercial licenses restrictions, and those resources which are widely accessible have been of poor quality, particularly with respect to reliability. Recent efforts to develop a highly reliable and publicly accessible alternative to the existing datasets were met with numerous challenges – not the least of which was filling the gaps in power transmission line voltage ratings. To address the line voltage rating problem, we developed and tested a basic methodology that fuses knowledge and techniques from power systems, geography, and machine learning domains. Specifically, we identified predictors of nominal voltage that could be extracted from aerial imagery and developed a tree-based classifier to classify nominal line voltage ratings. Overall, we found that line support height, support span, and conductor spacing are the best predictors of voltage ratings, and that the classifier built with these predictors had a reliable predictive accuracy (that is, within one voltage class for four out of the five classes sampled). We applied our approach to a study area in Minnesota.     

The effects of visual realism,spatial abilities,and competition on performance in map-based route learning in men

We report on how visual realism might influence map-based route learning performance in a controlled laboratory experiment with 104 male participants in a competitive context. Using animations of a dot moving through routes of interest, we find that participants recall the routes more accurately with abstract road maps than with more realistic satellite maps. We also find that, irrespective of visual realism, participants with higher spatial abilities (high-spatial participants) are more accurate in memorizing map-based routes than participants with lower spatial abilities (low-spatial participants). On the other hand, added visual realism limits high-spatial participants in their route recall speed, while it seems not to influence the recall speed of low-spatial participants. Competition affects participants’ overall confidence positively, but does not affect their route recall performance neither in terms of accuracy nor speed. With this study, we provide further empirical evidence demonstrating that it is important to choose the appropriate map type considering task characteristics and spatial abilities. While satellite maps might be perceived as more fun to use, or visually more attractive than road maps, they also require more cognitive resources for many map-based tasks, which is true even for high-spatial users.     

磁法三维建模在龙王沟磁铁矿勘探中的应用

以龙王沟磁铁矿高精度磁测数据为基础,采用小波多尺度分解方法,对化极后的磁法数据进行多尺度分解,通过功率谱法计算小波多尺度分解后的各阶细节异常的场源深度。结合地形、地质及钻孔资料,采用2.5D人机交互反演对磁测剖面上的磁铁矿体进行拟合,利用Golden Software公司开发的Voxler软件,在虚拟三维环境下实现二维小波多尺度分解数据的可视化,建立三维地质体模型,直观的预测磁铁矿体的空间范围和分布形态,实现地质、物探和地上、地下信息一体化集成展示,进行专业分析评价和辅助决策,可以有效的提高地质找矿效果。     

2011~2020年勘查地球化学数据处理研究进展

我国建立了包含海量数据的高质量的勘查地球化学数据库,为矿产勘查、环境评价和地质调查等提供了重要的数据支撑。如何高效处理勘查地球化学数据,并从中发掘和识别深层次信息一直是勘查地球化学学科研究的热点和前沿领域。本文在系统调研国内外学者过去十年发表的论著基础上,对勘查地球化学数据处理方法进行分析与对比,从勘查地球化学数据库建设、地球化学异常识别及其不确定性评价等方面概述了我国近十年来在该领域取得的主要研究进展,包括:(1)分形与多重分形模型由于考虑了地球化学空间模式的复杂性和尺度不变性,在全球范围内得到极大的发展和推广,我国学者引领了基于分形与多重分形的勘查地球化学数据处理;(2)机器学习和大数据思维开始在该领域启蒙,并迅速得到关注,正在成为研究热点和前沿领域,我国学者率先开展基于机器学习算法的勘查地球化学大数据挖掘研究;(3)我国学者需要进一步加强勘查地球化学数据缺失值处理以及成分数据闭合效应研究。今后该领域应进一步加强对弱缓地球化学异常识别、异常不确定性评价以及异常识别与其形成机理相结合等方面的研究。     

Spatial landslide susceptibility assessment using machine learning techniques assisted by additional data created with generative adversarial networks

In recent years,landslide susceptibility mapping has substantially improved with advances in machine learning.However,there are still challenges remain in landslide mapping due to the availability of limited inventory data.In this paper,a novel method that improves the performance of machine learning techniques is presented.The proposed method creates synthetic inventory data using Generative Adversarial Networks(GANs)for improving the prediction of landslides.In this research,landslide inventory data of 156 landslide locations were identified in Cameron Highlands,Malaysia,taken from previous projects the authors worked on.Elevation,slope,aspect,plan curvature,profile curvature,total curvature,lithology,land use and land cover(LULC),distance to the road,distance to the river,stream power index(SPI),sediment transport index(STI),terrain roughness index(TRI),topographic wetness index(TWI)and vegetation density are geo-environmental factors considered in this study based on suggestions from previous works on Cameron Highlands.To show the capability of GANs in improving landslide prediction models,this study tests the proposed GAN model with benchmark models namely Artificial Neural Network(ANN),Support Vector Machine(SVM),Decision Trees(DT),Random Forest(RF)and Bagging ensemble models with ANN and SVM models.These models were validated using the area under the receiver operating characteristic curve(AUROC).The DT,RF,SVM,ANN and Bagging ensemble could achieve the AUROC values of(0.90,0.94,0.86,0.69 and 0.82)for the training;and the AUROC of(0.76,0.81,0.85,0.72 and 0.75)for the test,subsequently.When using additional samples,the same models achieved the AUROC values of(0.92,0.94,0.88,0.75 and 0.84)for the training and(0.78,0.82,0.82,0.78 and 0.80)for the test,respectively.Using the additional samples improved the test accuracy of all the models except SVM.As a result,in data-scarce environments,this research showed that utilizing GANs to generate supplementary samples is promising because it can improve the predictive capability of common landslide prediction models.     

APG:A novel python-based ArcGIS toolbox to generate absence-datasets for geospatial studies

One important step in binary modeling of environmental problems is the generation of absence-datasets that are traditionally generated by random sampling and can undermine the quality of outputs.To solve this problem,this study develops the Absence Point Generation(APG)toolbox which is a Python-based ArcGIS toolbox for automated construction of absence-datasets for geospatial studies.The APG employs a frequency ratio analysis of four commonly used and important driving factors such as altitude,slope degree,topographic wetness index,and distance from rivers,and considers the presence locations buffer and density layers to define the low potential or susceptibility zones where absence-datasets are gener-ated.To test the APG toolbox,we applied two benchmark algorithms of random forest(RF)and boosted regression trees(BRT)in a case study to investigate groundwater potential using three absence datasets i.e.,the APG,random,and selection of absence samples(SAS)toolbox.The BRT-APG and RF-APG had the area under receiver operating curve(AUC)values of 0.947 and 0.942,while BRT and RF had weaker per-formances with the SAS and Random datasets.This effect resulted in AUC improvements for BRT and RF by 7.2,and 9.7％from the Random dataset,and AUC improvements for BRT and RF by 6.1,and 5.4％from the SAS dataset,respectively.The APG also impacted the importance of the input factors and the pattern of the groundwater potential maps,which proves the importance of absence points in environmental bin-ary issues.The proposed APG toolbox could be easily applied in other environmental hazards such as landslides,floods,and gully erosion,and land subsidence.