基于综合特征的纹理分割方法

针对灰度共生矩阵只能在单一尺度对纹理进行分析的不足,结合离散框架小波变换产生的尺度共生矩阵与梯度变换图像的灰度共生矩阵,提出了一种具有多尺度分析特性的综合纹理特征提取算法,并利用该特征对纹理图像进行分割.仿真实验结果表明:与基于单一尺度特征的纹理分割方法相比,本文提出的算法能够提高纹理边界定位准确性,减少区域内像素错分,取得了较好的分割效果.     

Mapping vegetation morphology types in a dry savanna ecosystem: integrating hierarchical object-based image analysis with Random Forest

Savanna ecosystems are geographically extensive and both ecologically and economically important, and require monitoring over large spatial extents. Remote-sensing-based characterization of vegetation properties in savannas is methodologically challenging, mainly due to high structural and functional heterogeneity. Recent advances in object-based image analysis (OBIA) and machine learning algorithms offer new opportunities to address these challenges. Focusing on the semi-arid savanna ecosystem in the central Kalahari, this study examined the suitability of a hierarchical OBIA approach combined with in situ data and an ensemble classification technique for mapping vegetation morphology types at landscape scale. A stack of Landsat TM imagery, NDVI, and topographic variables was segmented with six different scale factors resulting in a hierarchical network of image objects. Sample objects for each vegetation morphology class were selected at each segmentation scale and classification was performed using optimal features consisting of spectral and textural features. Overall and class-specific classification accuracies were compared across the six scales to examine the influence of segmentation scale on each. Results suggest that the highest overall classification accuracy (i.e. 85.59%) was observed not at the finest segmentation scale, but at coarse segmentation. Additionally, individual vegetation morphology classes differed in the segmentation scale at which they achieved highest classification accuracy, reflecting their unique ecology and physiognomic composition. While classes with high vegetation density/height attained higher accuracy at fine segmentation scale, those with lower vegetation density/height reached higher classification accuracy at coarse segmentation scales. Contrarily, for pans and bare areas, accuracy was relatively unaffected by changing segmentation scale. Variable importance plots suggested that spectral features were the most important, followed by textural variables. These results show the utility of the OBIA approach and emphasize the requirement of multi-scale analysis for accurately characterizing savanna systems.     

Scale computation on high spatial resolution remotely sensed imagery multi-scale segmentation

Scale computation for multi-scale image segmentation is an important research area in geographic object-based image analysis (GEOBIA). For the highly spectral heterogeneity of high spatial resolution remotely sensed imagery (HSRRSI), and the changing sizes of geographic features and their spatially distributive patterns, it is difficult to build a global or local-scale calculation parameter model to effectively guide multi-scale segmentation parameters setting in large-scale regions. Usually, the segmentation parameters are used to measure the heterogeneous and homogenous adjacent pixels in spatial and spectral spaces simultaneously. It has been proved that the adaptive acquiring parameter of scale plays a key role in gaining precise segmentation results, and later it deeply influences the automatic recognition and post-processing of the physical image parcels (PIPs). However, in most cases, scale computation techniques still fail to guide segmentation to produce appropriate or repeatable results which should meet the practical production standard of GIS data based on GEOBIA. These techniques have not been summarized and classified and there is no review focusing on scale computation for HSRRSI multi-scale segmentation. We provide an overview of the state-of-the-art segmentation scale computation techniques which are mainly based on the spectral statistics and geometric characteristics, etc. Moreover, the pedigree of segmentation scale has been first time proposed, and the overall performance of each category is analysed. Especially, the methods of local variance, semivariance, and synthetic semivariance are presented. Then, the scale object selection (SOS) algorithm, spectral angle algorithm, and the RMAS (ratio of mean difference to neighbours (ABS) to standard deviation) are discussed at spectral domain. In addition, miscellaneous scale computation approaches are recognized as the important researching aspect. In order to clearly describe the scale computation on multi-scale image segmentation, we have proposed the new conceptions of semantic image object (SIO), PIP, particular scale of interest, symbiotic scale, etc. At last, the trends of scale computation for HSRRSI multi-scale segmentation also have been presented.     

Analysis of co‐occurrence and discrete wavelet transform textures for differentiation of forest and non‐forest vegetation in very‐high‐resolution optical‐sensor imagery

The extraction of texture features from high‐resolution remote sensing imagery provides a complementary source of data for those applications in which the spectral information is not sufficient for identification or classification of spectrally similar landscape features. This study presents the results of grey‐level co‐occurrence matrix (GLCM) and wavelet transform (WT) texture analysis for forest and non‐forest vegetation types differentiation in QuickBird imagery. Using semivariogram fitting, the optimal GLCM windows for the land cover classes within the scene were determined. These optimal window sizes were then applied to eight GLCM texture measures (mean, variance, homogeneity, dissimilarity, contrast, entropy, angular second moment, and correlation) for the scene classification. Using wavelet transformation, up to five levels of macro‐texture were computed and tested in the classification process. Comparing the classification results, (1) the spectral‐only bands classification gave an overall accuracy of 58.69%; (2) the statistically derived 21×21 optimal mean texture combined with spectral information gave the best results among the GLCM optimal windows with an accuracy of 73.70%; and (3) the combined optimal WT‐texture levels 4 and 5 gave an accuracy of 63.56%. The combined classification of these three optimal results gave an overall accuracy of 77.93%. The results indicate that even though vegetation texture was generally measured better by the GLCM‐mean texture (micro‐textures) than by WT‐derived texture (macro‐textures), the results show that the micro–macro texture combination would improve the differentiation and classification of the overall vegetation types. Overall, the results suggests that computer‐assisted classification of high‐spatial‐resolution remotely sensed imagery has a good potential to augment the present ground‐based forest inventory methods.     

Practical guidelines for choosing GLCM textures to use in landscape classification tasks over a range of moderate spatial scales

Texture measurements quantitatively describe relationships of DN values of neighbouring pixels. The output is a continuous measure of spatial information that may be used for further processing. Spatial relationships are not necessarily correlated with spectral data for a given class, and including a measure of them improves classification accuracy. This research develops a guideline for choosing among the Haralick (Grey Level Co-occurrence Matrix [GLCM]) set of texture measures. These guidelines are derived using a variety of land covers and spatial scales (window sizes).

Principal component analysis (PCA) of eight GLCM measures was performed for three Landsat TM and ETM+ images: a mid-latitude agricultural and natural vegetation scene, a glacier–rock–sea ice scene, and a desert scene with dunes and structurally complex rocks. PCA was performed separately for neighbourhoods consisting of squares with 25, 13, and 5 pixels on a side to demonstrate robustness to different spatial scales. PCA loadings show that contrast (Con), dissimilarity, entropy (Ent), and GLCM variance are most commonly associated with visual edges of land-cover patches; homogeneity, GLCM mean, GLCM correlation (GLCM Cor), and angular second moment are associated with patch interiors. Edge-highlighting textures account for most dataset variance but fail to differentiate among classes. Eigenchannels highlighting patch interior characteristics rely on GLCM mean and to some extent GLCM Cor. These two textures do contribute to distinguishing individual class signatures for classification purposes. Ent does not appear consistently in edge or interior groupings. Ent is interpreted as important to the textures of particular classes, but which classes is not generalized from one scene to another. Con is effective for outlining patch edges and may serve for object formation in geographic object-based image analysis (GEOBIA).

For classification purposes, the proposed guideline is a choose Mean and, where a class patch is likely to contain edge-like features within it, Con. Cor is an alternative for Mean in these situations, Dis may similarly be used in place of Con. For more detailed texture study, add Ent. This guideline does not constitute a complete texture analysis but may allow confident use of GLCM texture to enhance the efficiency of Landsat-based classification.     

Extracting multi-features and optimizing feature space with sparse auto-encoder over WorldView-2 images

Feature selection of very high-resolution (VHR) images is a key prerequisite for supervised classification. However, it is always difficult to acquire the features which have the highest correlation to the type of land cover for improving classification accuracy. To address this problem, this paper proposed a methodology of feature selection using the results of multiple segmentation via genetic algorithm (GA) and correlation feature selection (CFS) integrating sparse auto-encoder (SAE). Firstly, 61 features, including spectral features and spatial features, are extracted from the results of multi-scale segmentation over a WorldView-2 image in Xicheng District, Beijing. Then, 40-dimensional features and 30-dimensional features are derived from the selection with GA+CFS and the optimization with SAE, respectively. Thirdly, the final classification is achieved by logistic regression (LR) based on different subsets of features extracted from the WorldView-2 image. It is found that the result of feature selection could contribute to increase in the intra-species separation and reduction in the inner-species variability. Adding extra lower-ranked features appeared to reduce the accuracy of classification. The results indicate that the overall classification accuracy with 30-dimensional features reached 87.56%, and increased 5.61% compared to the results with 61-dimensional features. For the two kinds of optimized features, the Z-test values are all greater than 1.96, which implied that feature dimensionality reduction and feature space optimization could significantly improve the accuracy of image land cover classification. The texture features in the wavelet domain are the most important features for the study area in the WorldView-2 image classification. Adding wavelet and the grey-level co-occurrence matrix (GLCM) information, especially for GLCM features in wavelet, appeared not to improve classification accuracy. The SAE-based method can produce feature subsets for improving mapping accuracy more efficiently.     

Analysis of wavelet packet and statistical textures for object-oriented classification of forest-agriculture ecotones using SPOT 5 imagery

Textural features of high-resolution remote sensing imagery are a powerful data source for improving classification accuracy because using only spectral information is not sufficient for the classification of objects with within-field spectral variability. This study presents the methods of using an object-oriented texture analysis algorithm for improving high-resolution remote sensing imagery classification, including wavelet packet transform texture analysis, the grey-level co-occurrence matrix (GLCM) and local spatial statistics. Wavelet packet transform texture analysis, with the method of optimization and selection of wavelet texture for feature extraction, is a good candidate for object-oriented classification. Feature optimization is used to reduce the data dimensions in combinations of textural sub-bands and spectral bands. The result of the classification accuracy assessment indicates the improvement of texture analysis for object-oriented classification in this study. Compared with the traditional method that uses only spectral bands, the combination of GLCM homogeneity and spectral bands increases the overall accuracy from 0.7431 to 0.9192. Furthermore, wavelet packet transform texture analysis is the optimal method, increasing the overall accuracy to 0.9216 using a smaller data dimension. Local spatial statistical measures also increase the classification total accuracy, but only from 0.7431 to 0.8088. This study demonstrates that wavelet packet and statistical textures can be used to improve object-oriented classification; specifically, the texture analysis based on the multiscale wavelet packet transform is optimal for increasing the classification accuracy using a smaller data dimension.     

Selection and Evaluation of the Optimal Scale in Multi-scale Segmentation of Remote Sensing Images

Multi-scale segmentation is the premise and key step of Object-Based Image Analysis (OBIA). The quality of multi-scale segmentation directly affects the accuracy of object-oriented classification. However, scale selection and evaluation remains a challenge in multi-scale segmentation. According to the fact that the optimal segmentation scale of the remote sensing image is closely related to the complexity of the objects of the image, a top-down method to select the optimal scale based on the complexity of segmented objects is proposed. In the top-down segmentation process, image features of each segmented object are extracted to construct the complexity function, and the optimal scale of each object is determined by setting a threshold value and iterating calculation. Then, the segmentation results with the best scale are obtained and applied to the ZY-3 satellite multispectral image and the GF-2 fusion image to obtain segmentation and classification results. Qualitative visual evaluation method, unsupervised evaluation method and supervised classification evaluation method were used to compare them with results obtained by the optimal single-scale segmentation and the unsupervised evaluation method. The experimental results show that the method can accurately obtain the scale matching with the ground targets, and improve segmentation effect and the classification accuracy, it is of practical value.     

遥感影像多尺度分割中最优尺度的选取及评价

多尺度分割是面向对象图像分析技术的前提和关键,多尺度分割的质量直接影响着面向对象分类的精度,但尺度选择仍然是多尺度分割中的一个难题。针对此问题,根据遥感影像的最优分割尺度与影像上目标复杂度密切相关的事实,提出了一种自上而下基于分割对象复杂度选取最优尺度的方法。该方法在分割过程中,提取每一对象的影像特征构建其复杂度函数,通过设置阈值,经迭代计算来确定每一对象的最优分割尺度,进而得到具有全局最优尺度的分割结果,并将其应用于ZY-3多光谱数据和GF-2融合影像,得到分割和分类结果。并将其与单一最优尺度和非监督评价法的分割及分类结果进行比较,结果表明：该方法能够获取与地面目标相匹配的分割尺度,改善了分割效果,提高了分类精度,具有一定实用价值。     

基于遥感光谱和空间变量随机森林的黄河三角洲刺槐林健康等级分类

对刺槐林健康状况进行准确分类制图,是进行刺槐林健康状况评估与生态修复的前提。以高分辨率IKONOS影像、基于影像提取的不同窗口、不同灰度共生矩阵纹理信息以及反映局部空间自相关的Local Getis-Ord Gi(Getis统计量)为数据源,结合实测生态样方数据,利用多决策树的组合分类模型随机森林(RF)对刺槐林健康进行分级,对6种方法的分类精度进行了比较且对分类变量的重要性进行了排序。结果显示:19m×19m是最佳纹理计算窗口;灰度共生矩阵均值是最优纹理变量;基于波段4计算的Getis统计量对RF分类具有最重要的作用;较之利用全部光谱、纹理和Getis统计量的80个波段/变量,利用前向选择得到的前16个重要性变量进行RF分类,获得了最高的分类精度(总精度为93.14%,Kappa系数为0.894)。研究证实了从高分影像提取的空间特征信息有助于提高对具有规则分布格局的人工刺槐林健康等级的分类精度;前向选择方法可以利用较少的预测变量获得较高的分类精度。     

基于模糊灰度共生矩阵与隐马尔可夫模型的断口图像识别

纹理通常由空间分布和灰度分布共同描述,灰度共生矩阵(GLCM)能兼顾二者,故广泛应用于纹理分析中。在计算GLCM时,为降低其维数,需对纹理图像进行灰度量化,这必然丢失部分图像信息。灰度量化时,由灰度值与量化区间中心值的不同距离,构造出相应的模糊隶属度函数,并定义了模糊灰度共生矩阵(FGLCM)。通过对断口图像FGLCM的14个特征统计量进行相关性分析,选择角二阶矩和熵等7个统计量作为特征参数,并验证了其有效性。最后,在4类典型断口图像的特征空间上,采用隐马尔可夫模型(HMM)进行分类识别。实践表明,FGLCM比已有的GLCM能更好地表征断口特性,且在HMM状态数为3时,断口分类的平均识别率可达98%。     

Improvement of classification in urban areas by the use of textural features: The case study of Lucknow city,Uttar Pradesh

Investigations have been carried out for digital spectral and textural classification of an Indian urban environment using SPOT images with grey level co-occurrence matrix (GLCM), grey level difference histogram (GLDH), and sum and difference histogram (SADH) approaches. The results indicate that a combination of texture and spectral features significantly improves the classification accuracy compared with classification with pure spectral features only. This improvement is about 9% and 17% for an addition of one and two texture features, respectively. GLDH and SADH give statistically similar results to GLCM, and take less computing time than GLCM. Conventional separability measures like transformed divergence, Bhattacharya distance, etc. are not effective in feature selection when classification is carried out with spectral and texture features. An alternative approach using simple statistics such as average coefficient of variation, skewness, and kurtosis and correlation amongst feature sets has shown greater feature selection potential when a combination of spectral and texture features is used.     

Multiscale Bayesian texture segmentation using neural networks and Markov random fields

This paper presents a wavelet-based texture segmentation method using multilayer perceptron (MLP) networks and Markov random fields (MRF) in a multi-scale Bayesian framework. Inputs and outputs of MLP networks are constructed to estimate a posterior probability. The multi-scale features produced by multi-level wavelet decompositions of textured images are classified at each scale by maximum a posterior (MAP) classification and the posterior probabilities from MLP networks. An MRF model is used in order to model the prior distribution of each texture class, and a factor, which fuses the classification information through scales and acts as a guide for the labeling decision, is incorporated into the MAP classification of each scale. By fusing the multi-scale MAP classifications sequentially from coarse to fine scales, our proposed method gets the final and improved segmentation result at the finest scale. In this fusion process, the MRF model serves as the smoothness constraint and the Gibbs sampler acts as the MAP classifier. Our texture segmentation method was applied to segmentation of gray-level textured images. The proposed segmentation method shows better performance than texture segmentation using the hidden Markov trees (HMT) model and the HMTseg algorithm, which is a multi-scale Bayesian image segmentation algorithm.     

基于注意力机制的多尺度融合航拍影像语义分割

航拍影像同一场景不同对象尺度差异较大,采用单一尺度的分割往往无法达到最 佳的分类效果。为解决这一问题,提出一种基于注意力机制的多尺度融合模型。首先,利用不 同采样率的扩张卷积提取航拍影像的多个尺度特征;然后,在多尺度融合阶段引入注意力机制, 使模型能够自动聚焦于合适的尺度,并为所有尺度及每个位置像素分别赋予权重;最后,将加 权融合后的特征图上采样到原图大小,对航拍影像的每个像素进行语义标注。实验结果表明, 与传统的 FCN、DeepLab 语义分割模型及其他航拍影像分割模型相比,基于注意力机制的多尺 度融合模型不仅具有更高的分割精度,而且可以通过对各尺度特征对应权重图的可视化,分析 不同尺度及位置像素的重要性。     

Classification of Remote Sensing Image based on Object-oriented Method:A Case Study of Baixiang County

Remote sensing is the main means of extracting land cover types,which has important significance for monitoring land use change and developing national policies.Object-based classification methods can provide higher accuracy data than pixel-based methods by using spectral,shape and texture information.In this study,we choose GF-1 satellite’s imagery and proposed a method which can automatically calculate the optimal segmentation scale.The object-based methods for classifying four typical land cover types are compared using multi-scale segmentation and three supervised machine learning algorithms.The relationship between the accuracy of classification results and the training sample proportion is analyzed and the result shows that object-based methods can achieve higher classification results in the case of small training sample ratio,overall accuracies are higher than 94%.Overall,the classification accuracy of support vector machine is higher than that of neural network and decision tree during the process of object-oriented classification.     

一种新颖的多尺度无监督图像分割算法

吉伯斯随机场(Gibbs Random Fields,GRF)作为一种引入图像空间信息的先验模型已广泛运用于贝叶斯图像分割中.然而迄今为止,所涉及的这类先验模型往往仅体现为单一尺度上的马尔科夫性,而在多尺度意义上却未曾涉及.首次通过扩展传统单尺度意义上GRF模型到多尺度上,即多尺度吉伯斯随机场,从而圆满地解决这些难题.实验表明:所提出的模型算法有很好的鲁棒性,且易于实现对图像无监督的精确分割.     

Study of urban spatial patterns from SPOT panchromatic imagery using textural analysis

The long-time historical evolution and recent rapid development of Beijing, China, present before us a unique urban structure. A 10-metre spatial resolution SPOT panchromatic image of Beijing has been studied to capture the spatial patterns of the city. Supervised image classifications were performed using statistical and structural texture features produced from the image. Textural features, including eight texture features from the Grey-Level Co-occurrence Matrix (GLCM) method; a computationally efficient texture feature, the Number of Different Grey-levels (NDG); and a structural texture feature, Edge Density (ED), were evaluated. It was found that generally single texture features performed poorly. Classification accuracy increased with increasing number of texture features until three or four texture features were combined. The more texture features in the combination, the smaller difference between different combinations. The results also show that a lower number of texture features were needed for more homogeneous areas. NDG and ED combined with GLCM texture features produced similar results as the same number of GLCM texture features. Two classification schemes were adopted, stratified classification and non-stratified classification. The best stratified classification result was better than the best non-stratified classification result.     

Does spatial resolution matter? A multi-scale comparison of object-based and pixel-based methods for detecting change associated with gas well drilling operations

An implicit assumption of the geographic object-based image analysis (GEOBIA) literature is that GEOBIA is more accurate than pixel-based methods for high spatial resolution image classification, but that the benefits of using GEOBIA are likely to be lower when moderate resolution data are employed. This study investigates this assumption within the context of a case study of mapping forest clearings associated with drilling for natural gas. The forest clearings varied from 0.2 to 9.2 ha, with an average size of 0.9 ha. National Aerial Imagery Program data from 2004 to 2010, with 1 m pixel size, were resampled through pixel aggregation to generate imagery with 2, 5, 15, and 30 m pixel sizes. The imagery for each date and at each of the five spatial resolutions was classified into Forest and Non-forest classes, using both maximum likelihood and GEOBIA. Change maps were generated through overlay of the classified images. Accuracy evaluation was carried out using a random sampling approach. The 1 m GEOBIA classification was found to be significantly more accurate than the GEOBIA and per-pixel classifications with either 15 or 30 m resolution. However, at any one particular pixel size (e.g. 1 m), the pixel-based classification was not statistically different from the GEOBIA classification. In addition, for the specific class of forest clearings, accuracy varied with the spatial resolution of the imagery. As the pixel size coarsened from 1 to 30 m, accuracy for the per-pixel method increased from 59% to 80%, but decreased from 71% to 58% for the GEOBIA classification. In summary, for studying the impact of forest clearing associated with gas extraction, GEOBIA is more accurate than pixel-based methods, but only at the very finest resolution of 1 m. For coarser spatial resolutions, per-pixel methods are not statistically different from GEOBIA.     

Exploring the synergistic use of multi-scale image object metrics for land-use/land-cover mapping using an object-based approach

This study presents a new method for the synergistic use of multi-scale image object metrics for land-use/land-cover mapping using an object-based classification approach. This new method can integrate an object with its super-objects’ metrics. The entire classification involves two object hierarchies: (1) a five-level object hierarchy to extract object metrics at five scales, and (2) a three-level object hierarchy for the classification process. A five-level object hierarchy was developed through multi-scale segmentation to calculate and extract both spectral and textural metrics. Layers representing the hierarchy at each of the five scales were then intersected by using the overlay tool, an intersected layer was created with metrics from all five scales, and the same geometric elements were retained as those of the objects of the lowest level. A decision tree analysis was then used to build rules for the classification of the intersected layer, which subsequently served as the thematic layer in a three-level object hierarchy to identify shadow regions and produce the final map. The use of multi-scale object metrics yielded improved classification results compared with single-scale metrics, which indicates that multi-scale object metrics provide valuable spatial information. This method can fully utilize metrics at multiple scales and shows promise for use in object-based classification approaches.     

地物提取的多尺度特征遥感应用分析

通过空间尺度效应分析,阐述不同属性景观地物在同一分辨率或同一尺度影像中提取的不合理性。为获得精确的地表信息,提出多尺度遥感影像分析方法,解决不同地物在不同空间尺度影像数据中提取的难题。通过多种分辨率影像的多尺度影像信息提取的应用实践,分析地物提取中的多尺度特性、尺度与分辨率关系等。