Margin-based feature selection for hyperspectral data

A margin-based feature selection approach is explored for hyperspectral data. This approach is based on measuring the confidence of a classifier when making predictions on a test data. Greedy feature flip and iterative search algorithms, which attempts to maximise the margin-based evaluation functions, were used in the present study. Evaluation functions use linear, zero–one and sigmoid utility functions where a utility function controls the contribution of each margin term to the overall score. The results obtained by margin-based feature selection technique were compared to a support vector machine-based recurring feature elimination approach. Two different hyperspectral data sets, one consisting of 65 bands (DAIS data) and other with 185 bands (AVIRIS data) were used. With digital airborne imaging spectrometer (DAIS) data, the classification accuracy by greedy feature flip algorithm and sigmoid utility function was 93.02% using a total of 24 selected features in comparison to an accuracy of 91.76% with full set of 65 features. The results suggest a significant increase in classification accuracy with 24 selected features. The classification accuracy (93.4%) achieved by the iterative search margin-based algorithm with 20 selected features using sigmoid utility function is also significantly more accurate than that achieved with 65 features. To judge the usefulness of margin-based feature selection approaches, another hyperspectral data set consisting of 185 features was used. A total of 65 selected features were used to evaluate the performance of margin-based feature selection approach. The results suggest a significantly improved performance by greedy feature flip-based feature selection technique with this data set also. This study also suggest that margin-based feature selection algorithms provide a comparable performance to support vector machine-based recurring feature elimination approach.     

Mapping the distributions of C3 and C4 grasses in the mixed-grass prairies of southwest Oklahoma using the Random Forest classification algorithm

The objective of this paper is to demonstrate a new method to map the distributions of C₃ and C₄ grasses at 30 m resolution and over a 25-year period of time (1988–2013) by combining the Random Forest (RF) classification algorithm and patch stable areas identified using the spatial pattern analysis software FRAGSTATS. Predictor variables for RF classifications consisted of ten spectral variables, four soil edaphic variables and three topographic variables. We provided a confidence score in terms of obtaining pure land cover at each pixel location by retrieving the classification tree votes. Classification accuracy assessments and predictor variable importance evaluations were conducted based on a repeated stratified sampling approach. Results show that patch stable areas obtained from larger patches are more appropriate to be used as sample data pools to train and validate RF classifiers for historical land cover mapping purposes and it is more reasonable to use patch stable areas as sample pools to map land cover in a year closer to the present rather than years further back in time. The percentage of obtained high confidence prediction pixels across the study area ranges from 71.18% in 1988 to 73.48% in 2013. The repeated stratified sampling approach is necessary in terms of reducing the positive bias in the estimated classification accuracy caused by the possible selections of training and validation pixels from the same patch stable areas. The RF classification algorithm was able to identify the important environmental factors affecting the distributions of C₃ and C₄ grasses in our study area such as elevation, soil pH, soil organic matter and soil texture.     

A new supervised classifier exploiting spectral-spatial information in the Bayesian framework

Conventional machine learning methods are often unable to achieve high degrees of accuracy when only spectral data are involved in the classification process. The main reason of that inaccuracy can be brought back to the omission of the spatial information in the classification. The present paper suggests a way to combine effectively the spectral and the spatial information and improve the classification’s accuracy. In practice, a Bayesian two-stage methodology is proposed embodying two enhancements: i) a geostatistical non-parametric classification approach, the universal indicator kriging and ii) the smooth multivariate kernel method. The former provides an informative prior, while the latter overcomes the assumption (often not true) of independence of the spectral data. The case study reports an application to land-cover classification in a study area located in the Apulia region (Southern Italy). The methodology performance in terms of overall accuracy was compared with five state-of-the-art methods, i.e. naïve Bayes, Random Forest, artificial neural networks, support vector machines and decision trees. It is shown that the proposed methodology outperforms all the compared methods and that even a severe reduction of the training set does not affect seriously the average accuracy of the presented method.     

An Evaluation of Four Different Wavelet Decomposition Procedures for Spatial Feature Discrimination in Urban Areas

Texture or spatial arrangement of neighborhood objects and features plays an important role in the human visual system for pattern recognition and image classification. The traditional spectral–based image processing techniques have proven inadequate for urban land use and land cover mapping from images acquired by the current generation of fine–resolution satellites. This is because of the high frequency spatial arrangements or complex nature of urban features. There is a need for an effective algorithm to digitally classify urban land use and land cover categories using high–resolution image data. Recent studies using wavelet transforms for texture analysis have generally reported better accuracy. Based on a high–resolution ATLAS image, this study illustrates four different wavelet decomposition procedures – the standard, horizontal, vertical, and diagonal decompositions – for urban land use and land cover feature extraction with the use of 33×33 pixel samples. The standard decomposition approach was found to be the most efficient approach in urban texture analysis and classification. For comparison purposes and to better evaluate the accuracy of wavelet approaches in image classification, spatial autocorrelation techniques (Moran's I and Geary's C ) and the spatial co–occurrence matrix method were also examined. The results suggest that the wavelet transform approach is superior to all other approaches.     

Mapping multiple tree species classes using a hierarchical procedure with optimized node variables and thresholds based on high spatial resolution satellite data

ABSTRACT

Tree species distribution mapping using remotely sensed data has long been an important research area. However, previous studies have rarely established a comprehensive and efficient classification procedure to obtain an accurate result. This study proposes a hierarchical classification procedure with optimized node variables and thresholds to classify tree species based on high spatial resolution satellite imagery. A classification tree structure consisting of parent and leaf nodes was designed based on user experience and visual interpretation. Spectral, textural, and topographic variables were extracted based on pre-segmented images. The random forest algorithm was used to select variables by ranking the impact of all variables. An iterating approach was used to optimize variables and thresholds in each loop by comprehensively considering the test accuracy and selected variables. The threshold range for each selected variable was determined by a statistical method considering the mean and standard deviation for two subnode types at each parent node. Classification of tree species was implemented using the optimized variables and thresholds. The results show that (1) the proposed procedure can accurately map the tree species distribution, with an overall accuracy of over 86% for both training and test stages; (2) critical variables for each class can be identified using this proposed procedure, and optimal variables of most tree plantation nodes are spectra related; (3) the overall forest classification accuracy using the proposed method is more accurate than that using the random forest (RF) and classification and regression tree (CART). The proposed approach provides results with 3.21% and 7.56% higher overall land cover classification accuracy and 4.68% and 10.28% higher overall forest classification accuracy than RF and CART, respectively.     

基于颜色特征利用色矩与BTC法进行影像聚类

影像聚类是一种对影像数据进行分组的方法,在基于内容的影像检索中,如果能够利用较低层次的可视特征进行高效的影像聚类,将会大大提高影像检索的精度.文章分别利用色矩法与分块截短编码(BTC)方法提取影像颜色特征,然后采用K均值聚类算法来对两种方法进行聚类分析.实验结果表明,分块截短编码(BTC)方法的聚类精度优于色矩法.     

Improving Spectral–spatial Classification of Hyperspectral Imagery Using Spectral Dimensionality Reduction Based on Weighted Genetic Algorithm

This paper aims to improve the accuracy and the efficiency of high resolution land cover mapping in urban area. To this end, an improved approach for classification of hyperspectral imagery is proposed and evaluated. This approach benefits from both inherent spectral and spatial information of an image. The weighted genetic (WG) algorithm is first used to obtain the subspace of hyperspectral data. The obtained features are then fed into the enhanced marker-based minimum spanning forest (EMSF) classification algorithm. In this algorithm, the markers are extracted from the classification maps obtained by both support vector machine and watershed segmentation algorithm classifiers. For this purpose, the class’s pixels with the largest population in the classification map are kept for each region of the segmentation map. Then, the most reliable classified pixels are chosen from among the exiting pixels as markers. To evaluate the efficiency of the proposed approach, three hyperspectral data sets acquired by ROSIS-03, Hymap and Hyper-Cam LWIR are used. Experimental results showed that the proposed WG–EMSF approach achieves approximately 9, 8 and 6% better overall accuracy than the original MSF-based algorithm for these data sets respectively.     

面向对象的无人机遥感影像岩溶湿地植被遥感识别

以广西桂林会仙喀斯特国家湿地公园为研究区,以无人机航摄影像为数据源,综合利用面向对象的影像分析技术、随机森林算法、阈值分类方法和Boruta全相关特征变量选择算法进行岩溶湿地植被的遥感识别。结果表明：针对不同特征变量对岩溶湿地遥感识别的贡献率而言,光谱特征（DOM > DSM） > 纹理特征（DOM > DSM） > 几何特征 > 上下文变量;两个航摄影像数据集的总体分类精度都在85%以上,Kappa系数也高于0.85。本文研究结果对基于高空间分辨率无人机可见光影像的岩溶湿地植被遥感识别在特征变量选择、分割参数选择及方法选择方面具有一定的借鉴意义。     

融和启发信息的蚁群特征选择方法

特征提取和选择是模式识别核心问题之一,它极大地影响着分类器的设计和性能,高维的特征选择更是一个NP难题。针对特征选择这一组合优化及多目标优化问题,本文提出了改进的融合启发信息ACO(Antcolony optimization)特征选择的新方法,该算法比不用启发信息的ACO方法能更好地找出代表问题空间的最优特征子集,降低分类系统的搜索空间,从而提高搜索效率。以航空纹理影像的特征选择和分类问题为例,利用原始蚂蚁算法和改进的蚂蚁算法选择的特征分别进行识别,结果证明该算法不仅能够比没有改进的蚂蚁找出有效特征集、降低图像特征空间维数、减少图像分类的工作量,而且提高了分类识别正确率。     

Effectiveness of SID as Spectral Similarity Measure to Develop Crop Spectra from Hyperspectral Image

The present study was undertaken with the objective to check effectiveness of spectral information divergence (SID) to develop spectra from image for crop classes based on spectral similarity with field spectra. In multispectral and hyperspectral remote sensing, classification of pixels is obtained by statistical comparison (by means of spectral similarity) of known field or library spectra to unknown image spectra. Though these algorithms are readily used, little emphasis has been placed on use of various spectral similarity measures to develop crop spectra from the image itself. Hence, in this study methodology suggested to develop spectra for crops based on SID. Absorption features are unique and distinct; hence, validation of the developed spectra is carried out using absorption features by comparing it with field spectra and finding average correlation coefficient r?=?0.982 and computed SID equivalent r?=?0.989. Effectiveness of developed spectra for image classification was computed by probability of spectral discrimination (PSD) and resulted in higher probability for the spectra developed based on SID. Image classification was carried out using field spectra and spectra assigned by SID. Overall classification accuracy of the image classified by field spectra is 78.30% and for the image classified by spectra assigned through SID-based approach is 91.82%. Z test shows that image classification carried out using spectra developed by SID is better than classification carried out using field spectra and significantly different. Validation by absorption features, effectiveness by PSD and higher classification accuracy show possibility of new approach for spectra development based on SID spectral similarity measure.     

基于多时相影像的农业作物非参数与概率分类

本文提出了一种新的结合多光谱和变化检测技术的多时相卫星数据集分类方法。该方法以数理统计中的最近邻法为基础,其目标函数是使得正确分类的平均概率得到最优化,即把每个分类类别看成同等重要。该新算法被应用于一个农业作物分类的研究区域,并利用覆盖该区的不同季节的SPOT和LANDSAT TM多时像影像。结果表明,与单时像影像相比,使用五个不同季节的多时像影像可以充分地提高分类精度。为了说明该方法在大尺度范围内的效果,本文选取瑞典道拉河流域作为研究区。由于不同地物的分布高度重叠,不可能得到像元水平上满意的分类精度。这就需要引进一种新的概念：像元概率分类法。基于像元的概率向量可用于判别传统分类法的可靠性并测量单个像元的不确定性（熵）。概率分类法同时提供了不同地物的面积的无偏估计,无论所感兴趣的区域的大小。这已经在不同特性的耕地试验点进行了检验。     

基于分类规则挖掘的遥感影像分类研究

分析了目前遥感影像的统计分类、神经网络分类及基于符号知识的逻辑推理分类方法的优缺点.以GIS为平台,构建了多源空间数据库,将数据挖掘的思想和方法引入遥感影像分类中,提出了面向分类规则挖掘的遥感影像分类框架.针对遥感光谱数据及其他空间数据的特点,定义了连续属性样本分类概念和分割点评价指标,提出了一种新的连续属性样本分类规则挖掘算法.选择一个试验区,采用该算法分别对遥感光谱数据、遥感光谱和DEM数据相结合的数据进行分类规则挖掘、遥感影像分类和分类精度比较.结果表明：（1）该算法具有较高的分类精度;（2）加入DEM等与分类相关的其他空间数据可以提高遥感影像的分类精度.通过挖掘分类规则进行遥感影像分类,扩展了基于知识的逻辑推理分类方法中知识获取渠道,提高了分类规则获取的智能化程度.新的连续属性样本分类规则挖掘算法,扩展了归纳学习算法对连续属性样本分类的适应性.     

SegOptim—A new R package for optimizing object-based image analyses of high-spatial resolution remotely-sensed data

Geographic Object-based Image Analysis (GEOBIA) is increasingly used to process high-spatial resolution imagery, with applications ranging from single species detection to habitat and land cover mapping. Image segmentation plays a key role in GEOBIA workflows, allowing to partition images into homogenous and mutually exclusive regions. Nonetheless, segmentation techniques require a robust parameterization to achieve the best results. Frequently, inappropriate parameterization leads to sub-optimal results and difficulties in comparing distinct methods.Here, we present an approach based on Genetic Algorithms (GA) to optimize image segmentation parameters by using the performance scores from object-based classification, thus allowing to assess the adequacy of a segmented image in relation to the classification problem. This approach was implemented in a new R package called SegOptim, in which several segmentation algorithms are interfaced, mostly from open-source software (GRASS GIS, Orfeo Toolbox, RSGISLib, SAGA GIS, TerraLib), but also from proprietary software (ESRI ArcGIS). SegOptim also provides access to several machine-learning classification algorithms currently available in R, including Gradient Boosted Modelling, Support Vector Machines, and Random Forest.We tested our approach using very-high to high spatial resolution images collected from an Unmanned Aerial Vehicle (0.03 – 0.10 m), WorldView-2 (2 m), RapidEye (5 m) and Sentinel-2 (10 – 20 m) in six different test sites located in northern Portugal with varying environmental conditions and for different purposes, including invasive species detection and land cover mapping. The results highlight the added value of our novel comparison of image segmentation and classification algorithms. Overall classification performances (assessed through cross-validation with the Kappa index) ranged from 0.85 to 1.00. Pilot-tests show that our GA-based approach is capable of providing sound results for optimizing the parameters of different segmentation algorithms, with benefits for classification accuracy and for comparison across techniques. We also verified that no particular combination of an image segmentation and a classification algorithm is suited for all the tasks/objectives. Consequently, it is crucial to compare and optimize available methods to understand which one is more suited for a certain objective.Our approach allows a closer integration between the segmentation and classification stages, which is of high importance for GEOBIA workflows. The results from our tests confirm that this integration has benefits for comparing and optimizing both processes. We discuss some limitations of the SegOptim approach (and potential solutions) as well as a future roadmap to expand its current functionalities.     

基于蚁群优化的特征选择新方法

利用蚁群优化算法解决特征选择问题，以获得能代表问题空间的较优特征子集，并能降低分类系统的搜索空间。以航空纹理影像的特征选择和分类问题为例，利用主分量变换和蚁群优化算法分别对原始纹理影像特征集合进行特征提取、选择和分类。结果表明，本文方法不仅能够降低图像特征空间维数，减少图像分类的工作量，而且还可以提高分类识别的正确率。     

Some issues in contextual fuzzy <Emphasis Type="Italic">c</Emphasis>-means classification of remotely sensed data for land cover mapping

Earlier for the hard classification techniques contextual information was used to improve classification accuracy. While modelling the spatial contextual information for hard classifiers using Markov Random Field it has been found that Metropolis algorithm is easier to program and it performs better in comparison to the Gibbs sampler. In the present study it has been found that incase of soft contextual classification Metropolis algorithm fails to sample from a random field efficiently and from the analysis it was found that Metropolis algorithm is not suitable for soft contextual classification due to the high dimensionality of the soft outputs.     

Using a stochastic gradient boosting algorithm to analyse the effectiveness of Landsat 8 data for montado land cover mapping: Application in southern Portugal

This study aims to develop and propose a methodological approach for montado ecosystem mapping using Landsat 8 multi-spectral data, vegetation indices, and the Stochastic Gradient Boosting (SGB) algorithm. Two Landsat 8 scenes (images from spring and summer 2014) of the same area in southern Portugal were acquired. Six vegetation indices were calculated for each scene: the Enhanced Vegetation Index (EVI), the Short-Wave Infrared Ratio (SWIR32), the Carotenoid Reflectance Index 1 (CRI1), the Green Chlorophyll Index (CI_green), the Normalised Multi-band Drought Index (NMDI), and the Soil-Adjusted Total Vegetation Index (SATVI). Based on this information, two datasets were prepared: (i) Dataset I only included multi-temporal Landsat 8 spectral bands (LS8), and (ii) Dataset II included the same information as Dataset I plus vegetation indices (LS8 + VIs). The integration of the vegetation indices into the classification scheme resulted in a significant improvement in the accuracy of Dataset II’s classifications when compared to Dataset I (McNemar test: Z-value = 4.50), leading to a difference of 4.90% in overall accuracy and 0.06 in the Kappa value. For the montado ecosystem, adding vegetation indices in the classification process showed a relevant increment in producer and user accuracies of 3.64% and 6.26%, respectively. By using the variable importance function from the SGB algorithm, it was found that the six most prominent variables (from a total of 24 tested variables) were the following: EVI_summer; CRI1_spring; SWIR32_spring; B6_summer; B5_summer; and CI_green_summer.     

Multispectral remote sensing for mapping grassland degradation using the key indicators of grass species and edaphic factors

Land degradation is believed to be one of the most severe and widespread environmental problems. In South Africa, large areas of land have been identified as degraded, as shown by the lower vegetation cover. One of the major causes of grassland degradation is change in plant species composition that leads to presence of unpalatable grass species. Some grass species have been successfully used as indicators of different levels of grassland degradation in the country. This paper, therefore explores the possibility of mapping grassland degradation in Cathedral Peak, South Africa, using indicators of grass species and edaphic factors. Multispectral SPOT 5 data were used to produce a grassland degradation map based on the spatial distribution of decreaser (Themeda triandra) and increaser (Hyparrhenia hirta) species. To improve mapping accuracy, soil samples were collected from each species site and analysed for nutrient content. A t-test and machine learning random forest classification algorithm were applied for variable selection and classification using SPOT 5 data and edaphic variables. Results indicated that the decreaser and increaser grass species can be mapped with modest accuracy using SPOT 5 data (overall accuracy of 75.30%, quantity disagreement = 2 and allocation disagreement = 23). The classification accuracy was improved to 88.60%, 1 and 11 for overall accuracy, quantity and allocation disagreements, respectively, when SPOT 5 bands and edaphic factors were combined. The study demonstrated that an approach based on the integration of multispectral data and edaphic variables, which increased the overall classification accuracy by about 13%, is a suitable when adopting remote sensing to monitor grassland degradation.     

分层神经网络分类算法

提高遥感图像分类精度一直是爱到普遍关注的焦点问题。近年来,人工神经网络技术和分 处理技术由于它们的许多优点受到广泛欢迎。本文把两种技术结合起来,提出了分层神经网络的概念,并基于此设计了一种分层神经网络分类算法。通过与最大似然法的对比实验表明,这种分层神经网络分类算法可以明显地提高分类精度,并对不规则分布的复杂数据具有很强的处理能力。     

Enhanced land use/cover classification of heterogeneous tropical landscapes using support vector machines and textural homogeneity

Land use/cover classification is a key research field in remote sensing and land change science as thematic maps derived from remotely sensed data have become the basis for analyzing many socio-ecological issues. However, land use/cover classification remains a difficult task and it is especially challenging in heterogeneous tropical landscapes where nonetheless such maps are of great importance. The present study aims at establishing an efficient classification approach to accurately map all broad land use/cover classes in a large, heterogeneous tropical area, as a basis for further studies (e.g., land use/cover change, deforestation and forest degradation). Specifically, we first compare the performance of parametric (maximum likelihood), non-parametric (k-nearest neighbor and four different support vector machines – SVM), and hybrid (unsupervised–supervised) classifiers, using hard and soft (fuzzy) accuracy assessments. We then assess, using the maximum likelihood algorithm, what textural indices from the gray-level co-occurrence matrix lead to greater classification improvements at the spatial resolution of Landsat imagery (30 m), and rank them accordingly. Finally, we use the textural index that provides the most accurate classification results to evaluate whether its usefulness varies significantly with the classifier used. We classified imagery corresponding to dry and wet seasons and found that SVM classifiers outperformed all the rest. We also found that the use of some textural indices, but particularly homogeneity and entropy, can significantly improve classifications. We focused on the use of the homogeneity index, which has so far been neglected in land use/cover classification efforts, and found that this index along with reflectance bands significantly increased the overall accuracy of all the classifiers, but particularly of SVM. We observed that improvements in producer's and user's accuracies through the inclusion of homogeneity were different depending on land use/cover classes. Early-growth/degraded forests, pastures, grasslands and savanna were the classes most improved, especially with the SVM radial basis function and SVM sigmoid classifiers, though with both classifiers all land use/cover classes were mapped with producer's and user's accuracies of ∼90%. Our classification approach seems very well suited to accurately map land use/cover of heterogeneous landscapes, thus having great potential to contribute to climate change mitigation schemes, conservation initiatives, and the design of management plans and rural development policies.