Evaluation of different machine learning methods for land cover mapping of a Mediterranean area using multi-seasonal Landsat images and Digital Terrain Models

Land cover monitoring using digital Earth data requires robust classification methods that allow the accurate mapping of complex land cover categories. This paper discusses the crucial issues related to the application of different up-to-date machine learning classifiers: classification trees (CT), artificial neural networks (ANN), support vector machines (SVM) and random forest (RF). The analysis of the statistical significance of the differences between the performance of these algorithms, as well as sensitivity to data set size reduction and noise were also analysed. Landsat-5 Thematic Mapper data captured in European spring and summer were used with auxiliary variables derived from a digital terrain model to classify 14 different land cover categories in south Spain. Overall, statistically similar accuracies of over 91% were obtained for ANN, SVM and RF. However, the findings of this study show differences in the accuracy of the classifiers, being RF the most accurate classifier with a very simple parameterization. SVM, followed by RF, was the most robust classifier to noise and data reduction. Significant differences in their performances were only reached for thresholds of noise and data reduction greater than 20% (noise, SVM) and 25% (noise, RF), and 80% (reduction, SVM) and 50% (reduction, RF), respectively.     

Evaluation of multisource data for glacier terrain mapping: a neural net approach

Spectrally similar nature of land covers in a glacierized terrain hampers their automated mapping from multispectral satellite data, which may be overcome by using multisource data. In the present study, an artificial neural network (ANN)-based information extraction approach was applied for mapping the Kolahoi glacier and adjoining areas, using Landsat TM (Thematic Mapper) data and several ancillary layers such as image transformations and topographic attributes. Results reveal that ANN (highest overall accuracy (OA): 83.74%) outperforms maximum likelihood classifier (highest OA: 66.90%) and the incorporation of ancillary data into the classification process significantly enhances the mapping accuracy (>9%), particularly the addition of Near Infrared Red/Short Wave Infrared (NIR/SWIR) data to the spectral data. A nine-band combination dataset (spectral data, slope, Red/NIR and decorrelation stretch) was found to be the best multisource dataset. Results of the Z-tests (at 95% confidence level) also corroborate and statistically validate the above findings.     

Multiple spectral similarity metrics for surface materials identification using hyperspectral data

Modern hyperspectral imaging and non-imaging spectroradiometer has the capability to acquire high-resolution spectral reflectance data required for surface materials identification and mapping. Spectral similarity metrics, due to their mathematical simplicity and insensitiveness to the number of reference labelled spectra, have been increasingly used for material mapping by labelling reflectance spectra in hyperspectral data labelling. For a particular hyperspectral data set, the accuracy of spectral labelling depends considerably upon the degree of unambiguous spectral matching achieved by the spectral similarity metric used. In this work, we propose a new methodology for quantifying spectral similarity for hyperspectral data labelling for surface materials identification. Developed adopting the multiple classifier system architecture, the proposed methodology unifies into a single framework the differential performances of eight different spectral similarity metrics for the quantification of spectral matching for surface materials. The proposed methodology has been implemented on two types of hyperspectral data viz. image (airborne hyperspectral images) and non-image (library spectra) for numerous surface materials identification. Further, the performance of the proposed methodology has been compared with the support vector machines (SVM) approach, and with all the base spectral similarity metrics. The results indicate that, for the hyperspectral images, the performance of the proposed methodology is comparable with that of the SVM. For the library spectra, the proposed methodology shows a consistently higher (increase of about 30% when compared to SVM) classification accuracy. The proposed methodology has the potential to serve as a general library search method for materials identification using hyperspectral data.     

土地覆盖制图：基于最优化遥感数据的支撑向量机分类

遥感数据具有在不同空间、光谱和时间尺度上获取地表测量信息的能力,使其成为获取土地覆盖信息的一个主要数据源。影像分类即把卫星影像上的相关像元划分给某类已知的土地覆盖类型的过程。支撑向量机（SVMs）是一种土地覆盖分类的新技术。三种常用的SVMs是：基于线性和多项式的SVM以及具有高斯核函数的SVM分类器,分类能否成功地应用有赖于其各自选择的最佳参数。但是海量的遥感数据使得这些参数的确定速度十分缓慢。本文研究了一种新的基于最优化遥感数据压缩技术的SVM分类方法。研究显示用于获取SVM参数的数据量能够在不影响土地覆盖的分类精度的前提下进行压缩。数据压缩成功的应用于多项式和高斯核函数的SVM分类,而线性SVM的分类精度却非常低。     

利用卷积神经网络的高光谱图像分类

针对高光谱图像分类中对光谱信息利用不足的问题,提出一种基于卷积神经网络在光谱域开展的分类算法。该算法通过构建五层网络结构,逐像素对光谱信息开展分析,将全光谱段集合作为输入,利用神经网络展开代价函数值的计算,实现对光谱特征的提取与分类。实验中采用三组高光谱遥感影像数据进行对比分析,以India Pines数据集为例,提出的基于卷积神经网络的分类方法的分类正确率达到90.16%,比RBF-SVM方法高出2.56%,相比三种传统的深度学习方法高出1%~3%,训练速度也较为理想。实验结果表明,本文所提出的算法充分利用了高光谱图像中逐像素点的光谱域信息,能够有效提高分类正确率。与传统学习算法相比,在较少训练样本的情况下,更能发挥其良好的分类性能。     

Classifier ensembles for land cover mapping using multitemporal SAR imagery

SAR data are almost independent from weather conditions, and thus are well suited for mapping of seasonally changing variables such as land cover. In regard to recent and upcoming missions, multitemporal and multi-frequency approaches become even more attractive. In the present study, classifier ensembles (i.e., boosted decision tree and random forests) are applied to multi-temporal C-band SAR data, from different study sites and years. A detailed accuracy assessment shows that classifier ensembles, in particularly random forests, outperform standard approaches like a single decision tree and a conventional maximum likelihood classifier by more than 10% independently from the site and year. They reach up to almost 84% of overall accuracy in rural areas with large plots. Visual interpretation confirms the statistical accuracy assessment and reveals that also typical random noise is considerably reduced. In addition the results demonstrate that random forests are less sensitive to the number of training samples and perform well even with only a small number. Random forests are computationally highly efficient and are hence considered very well suited for land cover classifications of future multifrequency and multitemporal stacks of SAR imagery.     

A comparative assessment between object and pixel-based classification approaches for land use/land cover mapping using SPOT 5 imagery

Abstract

Land use/land cover (LULC) classification with high accuracy is necessary, especially in eco-environment research, urban planning, vegetation condition study and soil management. Over the last decade a number of classification algorithms have been developed for the analysis of remotely sensed data. The most notable algorithms are the object-oriented K-Nearest Neighbour (K-NN), Support Vector Machines (SVMs) and the Decision Trees (DTs) amongst many others. In this study, LULC types of Selangor area were analyzed on the basis of the classification results acquired using the pixel-based and object-based image analysis approaches. SPOT 5 satellite images with four spectral bands from 2003 and 2010 were used to carry out the image classification and ground truth data were collected from Google Earth and field trips. In pixel-based image analysis, a supervised classification was performed using the DT classifier. On the other hand, object-oriented (K-NN) image analysis was evaluated using standard nearest neighbour as classifier. Subsequently SVM object-based classification was performed. Five LULC categories were extracted and the results were compared between them. The overall classification accuracies for 2003 and 2010 showed that the object-oriented (K-NN) (90.5% and 91%) performed better results than the pixel-based DT (68.6% and 68.4%) and object-based SVM (80.6% and 78.15%). In general, the object-oriented (K-NN) performed better than both DTs and SVMs. The obtained LULC classification maps can be used to improve various applications such as change detection, urban design, environmental management and zooning.     

Evaluation of diverse classification approaches for land use/cover mapping in a Mediterranean region utilizing Hyperion data

Information on Earth's land surface cover is commonly obtained through digital image analysis of data acquired from remote sensing sensors. In this study, we evaluated the use of diverse classification techniques in discriminating land use/cover types in a typical Mediterranean setting using Hyperion imagery. For this purpose, the spectral angle mapper (SAM), the object-based and the non-linear spectral unmixing based on artificial neural networks (ANNs) techniques were applied. A further objective had been to investigate the effect of two approaches for training sites selection in the SAM classification, namely of the pixel purity index (PPI) and of the direct selection of training points from the Hyperion imagery assisted by a QuickBird imagery and field-based training sites. Object-based classification outperformed the other techniques with an overall accuracy of 83%. Sub-pixel classification based on the ANN showed an overall accuracy of 52%, very close to that of SAM (48%). SAM applied using the training sites selected directly from the Hyperion imagery supported by the QuickBird image and the field visits returned an increase accuracy by 16%. Yet, all techniques appeared to suffer from the relatively low spatial resolution of the Hyperion imagery, which affected the spectral separation among the land use/cover classes.     

机载LiDAR和高光谱融合实现温带天然林树种识别

将机载LiDAR(Light Detection and Ranging)与高光谱CASI(Compact Airborne Spectrographic Imager)数据融合,充分利用垂直结构信息和光谱信息进行温带森林树种分类,并与仅用高光谱数据的分类结果相比较,评估融合数据的树种分类能力。结合样地实测数据,首先用LiDAR获得的3维垂直结构信息对CASI影像上的林间空隙进行掩膜,提取林木冠层子集;然后对冠层子集分层掩膜,利用光谱曲线的一阶微分及曲线匹配技术,实现各树种训练样本的自动提取;利用SVM分类器对两种数据分类并比较精度。结果表明,融合数据的树种分类总体精度和Kappa系数(83.88%,0.80)优于仅使用CASI数据(76.71%、0.71),优势树种的制图精度为78.43%—89.22%,用户精度为75.15%—95.65%,整体也优于仅使用CASI的制图精度(68.51%—84.69%)和用户精度(63.34%—95.45%)。结果表明,机载LiDAR与CASI基于像元的融合对温带森林树种识别的精度较仅高光谱数据有较大提高。     

城市土地利用变化的不透水面覆盖度检测方法

通过分析城市中不透水面数量和分布的变化与城市土地利用变化之间的对应关系, 综合中、高分辨率遥感数据各自的优势, 运用CART算法进行城市不透水面覆盖度(ISP)遥感估算, 基于ISP制图结果对城市土地利用变化进行检测。以山东省泰安市为例开展实验研究, 结果表明, 与传统的变化检测方法相比, 基于ISP的变化检测方法, 不仅能够反映土地利用类型转换的潜在信息, 而且可以灵活地量化定义和解释城市用地变化情况。这种方法为城市土地利用变化信息的提取和分析提供了一种新的思路, 可以作为现有变化检测方法的有益补充。     

大坝变形预测的支持向量机模型

针对大坝变形具有强非线性的特点以及在采用传统神经网络模型进行预测时存在局部极小、过学习等问题,提出一种新的大坝变形预测方法——支持向量机方法。该方法基于统计学习理论,采用结构风险最小化原则,保证了模型具有很强的泛化性能,并通过求解一个二次规划问题确保模型具有全局最优。以东江大坝变形预测为实例,说明了该方法的可行性和有效性。     

图像分类中基于核映射的光谱匹配度量方法

针对多光谱遥感数据特点利用SSV匹配技术改进高斯核函数得到新的KSSV函数,然后在由KSSV核函数映射得到的高维空间中利用SAM匹配技术代替基于欧氏距离的相似性度量。如此可以充分挖掘多光谱影像中的波谱特征信息并有效利用,提高模式识别方法应用的有效性。将此方法分别应用于非监督分类(k均值)与监督分类(最小距离、SVM)的试验表明,改进度量的分类方法可显著提高地类间的可区分度并有效降低类内的不一致性,更有效针对多光谱遥感影像中的地物类型,获得较好的精度改进。     

入侵种互花米草的光谱分层分析方法

针对互花米草的爆发式增长对沿海滩涂生物多样性和生态稳定带来了巨大的生态威胁的问题,该文以闽江河口互花米草和其他3种湿地植物的室内叶片高光谱数据为例,探讨互花米草与其伴生植物是否具有光谱可分性。采用一种分层分析方法对实测高光谱数据降维并选择出识别互花米草的最佳波段。首先,利用ANOVA对光谱数据降维,选择出互花米草与其他湿地植物光谱具有显著性差异的波段;其次,使用CART算法对ANOVA降维后具有显著差异的高光谱数据进一步降维,找到识别互花米草潜在的最佳波段;最后,利用J-M距离评估CART选择波段的可分性。结果表明:互花米草与其他3种湿地植物具有光谱可分性,其J-M距离均高于1.9;基于CART算法的入侵种互花米草的识别精度平均达到96.7%,高于传统方法的识别精度。该文成果将为航空或航天高光谱遥感监测互花米草入侵区提供参考。     

基于特征空间中类间可分性的层次型多类支撑向量机

针对支撑向量机的特点提出了一种特征空间中的类间可分性度量,并基于该度量通过聚类算法构造了二叉树和单层聚类两种层次型多类支撑向量机。通过多光谱遥感影像的分类实验证明了该可分性度量的有效性。     

Assessment of Sentinel-1A data for rice crop classification using random forests and support vector machines

This study developed an approach to map rice-cropping systems in An Giang and Dong Thap provinces, South Vietnam using multi-temporal Sentinel-1A (S1A) data. The data were processed through four steps: (1) data pre-processing, (2) constructing smooth time series VH backscatter data, (3) rice crop classification using random forests (RF) and support vector machines (SVM) and (4) accuracy assessment. The results indicated that the smooth VH backscatter profiles reflected the temporal characteristics of rice-cropping patterns in the study region. The comparisons between the classification results and the ground reference data indicated that the overall accuracy and Kappa coefficient achieved from RF were 86.1% and 0.72, respectively, which were slightly more accurate than SVM (overall accuracy of 83.4% and Kappa coefficient of 0.67). These results were reaffirmed by the government’s rice area statistics with the relative error in area (REA) values of 0.2 and 2.2% for RF and SVM, respectively.     

A support vector machine classifier based on a new kernel function model for hyperspectral data

The kernel function is a key factor to determine the performance of a support vector machine (SVM) classifier. Choosing and constructing appropriate kernel function models has been a hot topic in SVM studies. But so far, its implementation can only rely on the experience and the specific sample characteristics without a unified pattern. Thus, this article explored the related theories and research findings of kernel functions, analyzed the classification characteristics of EO-1 Hyperion hyperspectral imagery, and combined a polynomial kernel function with a radial basis kernel function to form a new kernel function model (PRBF). Then, a hyperspectral remote sensing imagery classifier was constructed based on the PRBF model, and a genetic algorithm (GA) was used to optimize the SVM parameters. On the basis of theoretical analysis, this article completed object classification experiments on the Hyperion hyperspectral imagery of experimental areas and verified the high classification accuracy of the model. The experimental results show that the effect of hyperspectral image classification based on this PRBF model is apparently better than the model established by a single global or local kernel function and thus can greatly improve the accuracy of object identification and classification. The highest overall classification accuracy and kappa coefficient reached 93.246% and 0.907, respectively, in all experiments.     

Satellite image analysis using crowdsourcing data for collaborative mapping: current and opportunities

ABSTRACT

Researchers are continually finding new applications of satellite images because of the growing number of high-resolution images with wide spatial coverage. However, the cost of these images is sometimes high, and their temporal resolution is relatively coarse. Crowdsourcing is an increasingly common source of data that takes advantage of local stakeholder knowledge and that provides a higher frequency of data. The complementarity of these two data sources suggests there is great potential for mutually beneficial integration. Unfortunately, there are still important gaps in crowdsourced satellite image analysis by means of crowdsourcing in areas such as land cover classification and emergency management. In this paper, we summarize recent efforts, and discuss the challenges and prospects of satellite image analysis for geospatial applications using crowdsourcing. Crowdsourcing can be used to improve satellite image analysis and satellite images can be used to organize crowdsourced efforts for collaborative mapping.     

面向数字孪生城市的智能化全息测绘

以大数据、物联网、人工智能、虚拟现实、云计算、智能驾驶等新技术为代表的信息化浪潮席卷全球,数字世界与物理世界正形成两大平行发展、相互作用的体系,数字孪生技术应运而生。随着物联网技术（IOT）的发展,数字孪生的理念被引入到智慧城市建设中来,深刻影响着城市规划、建设与治理。笔者所在单位面向数字孪生城市和自然资源统一监管对测绘地理信息的新需求,在全国开创性地开展了面向数字孪生城市的智能化全息测绘试点工作。本文结合上海市智能化全息测绘试点工作,从数字孪生城市、数字孪生城市对地理信息的新需求、智能化全息测绘关键技术及测绘成果等方面展开了论述,重点介绍了智能化全息测绘的技术体系和产品体系,以及在社会各领域的应用成果。     

A comparative analysis of a fixed thresholding vs. a classification tree approach for operational burn scar detection and mapping

The scope of this paper is to demonstrate, evaluate and compare two burn scar mapping (BSM) approaches developed and applied operationally in the framework of the RISK-EOS service element project within the Global Monitoring for Environment and Security (GMES) program funded by ESA (http://www.risk-eos.com). The first method is the BSM_NOA, a fixed thresholding method using a set of specifically designed and combined image enhancements, whilst the second one is the BSM_ITF, a decision tree classification approach based on a wide range of biophysical parameters. The two methods were deployed and compared in the framework of operational mapping conditions set by RISK-EOS standards, based either on sets of uni- or multi-temporal satellite images acquired by Landsat 5 TM and SPOT 4 HRV. The evaluation of the performance of the two methods showed that either in uni- or multi-temporal acquisition mode, the two methods reach high detection capability rates ranging from 80% to 91%. At the same time, the minimum burnt area detected was of 0.9–1.0 ha, despite the coarser spatial resolution of Landsat 5 TM sensor. Among the advantages of the satellite-based approaches compared to conventional burn scar mapping, are cost-efficiency, repeatability, flexibility, and high spatial and thematic accuracy from local to country level. Following the catastrophic fire season of 2007, burn scar maps were generated using BSM_NOA for the entirety of Greece and BSM_ITF for south France in the framework of the RISK-EOS/GMES Services Element project.     

A GIS-based comparative study of Dempster-Shafer,logistic regression and artificial neural network models for landslide susceptibility mapping

The main aim of present study is to compare three GIS-based models, namely Dempster–Shafer (DS), logistic regression (LR) and artificial neural network (ANN) models for landslide susceptibility mapping in the Shangzhou District of Shangluo City, Shaanxi Province, China. At First, landslide locations were identified by aerial photographs and supported by field surveys, and a total of 145 landslide locations were mapped in the study area. Subsequently, the landslide inventory was randomly divided into two parts (70/30) using Hawths Tools in ArcGIS 10.0 for training and validation purposes, respectively. In the present study, 14 landslide conditioning factors such as altitude, slope angle, slope aspect, topographic wetness index, sediment transport index, stream power index, plan curvature, profile curvature, lithology, rainfall, distance to rivers, distance to roads, distance to faults and normalized different vegetation index were used to detect the most susceptible areas. In the next step, landslide susceptible areas were mapped using the DS, LR and ANN models based on landslide conditioning factors. Finally, the accuracies of the landslide susceptibility maps produced from the three models were verified using the area under the curve (AUC). The validation results showed that the landslide susceptibility map generated by the ANN model has the highest training accuracy (73.19%), followed by the LR model (71.37%), and the DS model (66.42%). Similarly, the AUC plot for prediction accuracy presents that ANN model has the highest accuracy (69.62%), followed by the LR model (68.94%), and the DS model (61.39%). According to the validation results of the AUC curves, the map produced by these models exhibits the satisfactory properties.