Pixel- and feature-level fusion of hyperspectral and lidar data for urban land-use classification

The complexity of urban areas makes it difficult for single-source remotely sensed data to meet all urban application requirements. Airborne light detection and ranging (lidar) can provide precise horizontal and vertical point cloud data, while hyperspectral images can provide hundreds of narrow spectral bands which are sensitive to subtle differences in surface materials. The main objectives of this study are to explore: (1) the performance of fused lidar and hyperspectral data for urban land-use classification, especially the contribution of lidar intensity and height information for land-use classification in shadow areas; and (2) the efficiency of combined pixel- and object-based classifiers for urban land-use classification. Support vector machine (SVM), maximum likelihood classification (MLC), and object-based classifiers were used to classify lidar, hyperspectral data and their derived features, such as the normalized digital surface model (nDSM), normalized difference vegetation index (NDVI), and texture measures, into 15 urban land-use classes. Spatial attributes and rules were used to minimize misclassification of the objects showing similar spectral properties, and accuracy assessments were carried out for the classification results. Compared with hyperspectral data alone, hyperspectral–lidar data fusion improved overall accuracy by 6.8% (from 81.7 to 88.5%) when the SVM classifier was used. Meanwhile, compared with SVM alone, the combined SVM and object-based method improved OA by 7.1% (from 87.6 to 94.7%). The results suggest that hyperspectral–lidar data fusion is effective for urban land-use classification, and the proposed combined pixel- and object-based classifiers are very efficient and flexible for the fusion of hyperspectral and lidar data.     

Improving detailed rule-based feature extraction of urban areas from WorldView-2 image and lidar data

Urbanization is commonly accepted as an important contributor to the growth of man-made structures and as a rapid convertor of natural environments to impervious surfaces. Roofs are one class of impervious surface whose materials can highly influence the quality of urban surface water. In this study, two data sources, WorldView-2 (WV-2) imagery and a combination of WV-2 and lidar data, were utilized to map intra-urban targets, with 13 classes. Images were classified using object-based image analysis. Pixel-based classifications using the support vector machine (SVM) and maximum likelihood (ML) methods were also tested for their abilities to use both lidar data and WV-2 imagery. ML and SVM classifications yielded overall accuracies of 72.46% and 75.69%, respectively. The results of these classifiers exhibited mixed pixels and salt-and-pepper effects. Spectral, spatial, and textural attributes as well as various spectral indices were employed in the object-based classification of WV-2 imagery. Feature classification of WV-2 imagery resulted in 85% overall accuracy. Lidar data were added to WV-2 imagery to assist in the spatial and spectral diversities of urban infrastructures. Classified image made from WV-2 imagery and lidar data achieved 92.84% overall accuracy. Rule-sets of these fused datasets effectively reduced the spectral variation and spatial heterogeneities of intra-urban classes, causing finer boundaries among land-cover classes. Therefore, object-based classification of WV-2 imagery and lidar data efficiently improved detailed characterization of roof types and other urban surface materials.     

Land cover classification in rugged areas using simulated moderateresolution remote sensor data and an artificial neural network

Rugged land cover classification accuracies produced by an artificial neural network (ANN) using simulated moderate-resolution remote sensor data exceed overall accuracies produced using the maximum likelihood rule (MLR). Land cover in spatially-complex areas and at broad spatial scales may be difficult to monitor due to ambiguities in spectral reflectance information produced from cloud-related and topographic effects, or from sampling constraints. Such ambiguities may produce inconsistent estimates of changes in vegetation status, surface energy balance, run-off yields, or other land cover characteristics. By use of a 'back-classification' protocol, which uses the same pixels for testing as for training the classifier, tests of ANN versus MLR-based classifiers demonstrated the ANNbased classifier equalled or exceeded classification accuracies produced by the MLR-based classifier in five of six land cover classes evaluated.     

On the classification of remote sensing high spatial resolution image data

In this paper is investigated a methodology implementing an object-based approach to digital image classification using spectral and spatial attributes in a multiple-stage classifier structured as a binary tree. It is a well-established fact that object-based image classification is particularly appropriate when dealing with high spatial resolution image data. Following this approach, the image is initially segmented into objects that carry informational value. Next, spectral and spatial attributes are extracted from every object in the scene, and implemented into a classifier to produce a thematic map. As the combined number of spectral and spatial variables may become large compared to the number of available training samples, a reduction in the data dimensionality may be required whenever parametric classifiers are used, in order to mitigate the effects of the Hughes phenomenon. To this end the sequential feature selection (SFS) procedure is applied in a multiple-stage classifier structured as a binary tree. The advantage of a binary tree classifier lies in the fact that only one pair of classes is considered at each stage (node), allowing for an optimal selection of features. This proposed approach was tested using Quickbird image data covering an urban scene. The results are compared against results yielded by the traditional single-stage Gaussian maximum likelihood classifier. The results suggest the proposed methodology is adequate in the classification of high spatial resolution image data.     

Mapping vegetation and land cover in a large urban area using a multiple classifier system

Vegetation and land-cover information is critical for sustainable environmental management in urban areas. Remote sensing has increasingly been used to derive such information, yet it has been challenged by the spectral and spatial complexity in the urban environment. In this study, we developed a multiple classifier system (MCS) to help improve remote-sensing-based vegetation and land-cover mapping in a large metropolitan area. MCSs, although considered as an emerging hot topic and a promising trend in pattern recognition, have not received the attention it deserves in the remote-sensing community. Our work consisted of several components. First, we identified a group of commonly used pattern recognizers from different families of statistical learning algorithms as base classifiers. Then, we implemented them to derive land-cover information from a satellite image covering the study site. Last, we adopted a weighting and combination method to generate the final map. Results indicate that there is statistically significant difference in the classification accuracy between the MCS developed and each base classifier considered. Comparing with the base classifiers, the MCS produced not only about 5–8% higher overall classification accuracy but also the most stable categorical accuracies. Moreover, the MCS generated a larger accuracy improvement for spectrally complex classes than for relatively homogenous ones, suggesting its comparative advantage in reducing classification errors caused by class ambiguity. The novelties of our work are with the demonstration of how MCSs can be operationally used to improve image classification from large remote sensor data sets with complex patterns and with the insight into the behaviour of MCSs in relation to the complexity of individual classes. These findings can help promote the use of MCSs as an emerging premier approach for image classification by the remote-sensing community.     

样本特征对参数/非参数分类器分类精度的影响分析

为验证理论训练数量(10~30 p)对参数分类器(如最大似然分类)、非参数分类器(如支撑向量机)的适用性以及样本特征(光谱统计、空间分布特征)对分类器分类精度的影响,选择不同规模的训练样本进行最大似然分类和支撑向量机分类,分析分类精度与样本之间的关系。实验结果表明:随着样本量的增加,最大似然、支撑向量机分类精度均随样本量增多而提高并趋于稳定,最大似然分类精度的增长速度要快于支撑向量机。MLC受样本量的影响较大,在小样本的时候(5个),分类精度不稳定,超过30个样本的时候,分类精度稳定下来;对于SVM分类器,在小样本的时候(5个),分类精度较高且稳定,因此SVM分类适合于小样本分类,不受限于理论样本量的影响。当样本量超过最小理论样本量值(30个)的时候,最大似然分类精度要优于支撑向量机,主要是由于当样本量增加后,最大似然更易于获得有效的信息量样本,而对于支撑向量机边缘信息样本的增加数量不大。研究结果为进一步优化样本进行分类打下前期的实验基础。     

Object-based classification using SPOT-5 imagery for Moso bamboo forest mapping

This study proposed a multi-scale, object-based classification analysis of SPOT-5 imagery to map Moso bamboo forest. A three-level hierarchical network of image objects was developed through multi-scale segmentation. By combining spectral and textural properties, both the classification tree and nearest neighbour classifiers were used to classify the image objects at Level 2 in the three-level object hierarchy. The feature selection results showed that most of the object features were related to the spectral properties for both the classification tree and nearest neighbour classifiers. Contextual information characterized by the composition of classified image objects using the class-related features assisted the detection of shadow areas at Levels 1 and 3. Better classification results were achieved using the nearest neighbour algorithm, with both the producer’s and user’s accuracy higher than 90% for Moso bamboo and an overall accuracy of over 85%. The object-based approach toward incorporating textural and contextual information in classification sequence at various scales shows promise in the analysis of forest ecosystems of a complex nature.     

The effects of combining classifiers with the same training statistics using Bayesian decision rules

Multiple classifier systems (MCS) are attracting increasing interest in the field of pattern recognition and machine learning. Recently, MCS are also being introduced in the remote sensing field where the importance of classifier diversity for image classification problems has not been examined. In this article, Satellite Pour l'Observation de la Terre (SPOT) IV panchromatic and multispectral satellite images are classified into six land cover classes using five base classifiers: contextual classifier, k-nearest neighbour classifier, Mahalanobis classifier, maximum likelihood classifier and minimum distance classifier. The five base classifiers are trained with the same feature sets throughout the experiments and a posteriori probability, derived from the confusion matrix of these base classifiers, is applied to five Bayesian decision rules (product rule, sum rule, maximum rule, minimum rule and median rule) for constructing different combinations of classifier ensembles. The performance of these classifier ensembles is evaluated for overall accuracy and kappa statistics. Three statistical tests, the McNemar's test, the Cochran's Q test and the Looney's F-test, are used to examine the diversity of the classification results of the base classifiers compared to the results of the classifier ensembles. The experimental comparison reveals that (a) significant diversity amongst the base classifiers cannot enhance the performance of classifier ensembles; (b) accuracy improvement of classifier ensembles can only be found by using base classifiers with similar and low accuracy; (c) increasing the number of base classifiers cannot improve the overall accuracy of the MCS and (d) none of the Bayesian decision rules outperforms the others.     

Land cover update by supervised classification of segmented ASTER images

The revision of the 1995 land cover dataset for the Vale do Sousa region, in the northwest of Portugal, was carried out by supervised classification of a multi‐spectral image from the Advanced Spaceborne Thermal Emission and Reflectance Radiometer (ASTER) sensor. The nine reflective bands of ASTER were used, covering the spectral range from 0.52–2.43?µm. The image was initially ortho‐rectified and segmented into 51?186 objects, with an average object size of 135?pixels (about 3?ha). A total of 582 of these objects were identified for training nine land cover classes. The image was classified using an algorithm based on a fuzzy classifier, Support Vector Machines (SVM), K Nearest Neighbours (K‐NN) and a Logistic Discrimination (LD) classifier. The results from the classification were evaluated using a set of 277 validation sites, independently gathered. The overall accuracy was 44.6% for the fuzzy classifier, 70.5% for the SVM, 60.9% for the K‐NN and 72.2% for the LD classifier. The difficulty in discriminating between some of the forest land cover classes was examined by separability analysis and unsupervised classification with hierarchical clustering. The forest classes were found to overlap in the multi‐spectral space defined by the nine ASTER bands used.     

结合面向对象方法的多层次干旱与半干旱区域城镇提取方法

在干旱与半干旱区域戈壁及沙漠等高亮地表与城镇连成一片,两者的光谱特征在中等分辨率遥感数据上非常相似;因此,利用基于像素的分类方法很难将城镇准确提取出来。根据两种地物的样本对NDVI、NDBI的分布特征统计分析得出:基于面向对象的分类方法在提取城镇信息方面有较大优势。以典型的干旱区域—黑河流域张掖市及周边地区作为研究区域,将面向对象的方法应用到具有中等分辨率的Landsat-TM数据上,提出了结合面向对象方法的多层次干旱与半干旱区域城镇提取方法。该方法首先使用分层分类的方法得到城镇和荒漠的混合影像,然后使用面向对象的分类方法精确提取城镇信息,其中分割对象过程中引入样本可分离度量化不同尺度的影像分割效果,实现最优尺度分割。结果表明:其目视效果、总体精度(94.51%)和Kappa系数(0.89),均优于支持向量机(SVM)与基于时间序列的分类方法。     

Land-cover classification from multiple classifiers using decision fusion based on the probabilistic graphical model

A novel method of using different classification algorithms in an integrated manner by adaptively weighted decision level fusion was proposed. The proposed fusion scheme involves two steps. First, we processed the data using each classifier separately and provided probability estimations for each pixel of the considered classes. Then, the results are aggregated on the basis of the decision rule of probabilistic graphical model according to the capabilities of classifiers and ancillary information. The method was tested and validated through the Landsat 8 operational land imager data using two different classifiers, namely, maximum likelihood classifier and support vector machine. The proposed method provided higher accuracy improvement than the separate use of different classifiers and that complex landscapes, such as mountainous regions, have higher accuracy improvement than the relatively homogenous ones. Moreover, the method can handle more than two types of classifiers and effectively introduce additional ancillary information for adaptive weight selection. These findings can help promote our proposed method as an emerging approach for land-cover classification through remote sensing technology.     

Improving global scale land cover classifications with multi-directional POLDER data and a decision tree classifier

Several investigations indicate that the Bidirectional Reflectance Distribution Function (BRDF) contains information that can be used to complement spectral information for improved land cover classification accuracies. Prior studies on the addition of BRDF information to improve land cover classifications have been conducted primarily at local or regional scales. Thus, the potential benefits of adding BRDF information to improve global to continental scale land cover classification have not yet been explored. Here we examine the impact of multidirectional global scale data from the first Polarization and Directionality of Earth Reflectances (POLDER) spacecraft instrument flown on the Advanced Earth Observing Satellite (ADEOS-1) platform on overall classification accuracy and per-class accuracies for 15 land cover categories specified by the International Geosphere Biosphere Programme (IGBP).

A set of 36,648 global training pixels (7 × 6 km spatial resolution) was used with a decision tree classifier to evaluate the performance of classifying POLDER data with and without the inclusion of BRDF information. BRDF ‘metrics’ for the eight-month POLDER on ADEOS-1 archive (10/1996–06/1997) were developed that describe the temporal evolution of the BRDF as captured by a semi-empirical BRDF model. The concept of BRDF ‘feature space’ is introduced and used to explore and exploit the bidirectional information content. The C5.0 decision tree classifier was applied with a boosting option, with the temporal metrics for spectral albedo as input for a first test, and with spectral albedo and BRDF metrics for a second test. Results were evaluated against 20 random subsets of the training data.

Examination of the BRDF feature space indicates that coarse scale BRDF coefficients from POLDER provide information on land cover that is different from the spectral and temporal information of the imagery. The contribution of BRDF information to reducing classification errors is also demonstrated: the addition of BRDF metrics reduces the mean, overall classification error rates by 3.15% (from 18.1% to 14.95% error) with larger improvements for producer's accuracies of individual classes such as Grasslands (+ 8.71%), Urban areas (+ 8.02%), and Wetlands (+ 7.82%). User's accuracies for the Urban (+ 7.42%) and Evergreen Broadleaf Forest (+ 6.70%) classes are also increased. The methodology and results are widely applicable to current multidirectional satellite data from the Multi-angle Imaging Spectroradiometer (MISR), and to the next generation of POLDER-like multi-directional instruments.     

Development of an IKONOS image classification rule-set for multi-scale mapping of Mediterranean rural landscapes

In this research, a rule-set of object-based classification of IKONOS imagery for fine-scale mapping of Mediterranean rural landscapes was developed. This study was conducted on the Mediterranean island of Crete (Greece). A three-level classification hierarchy was designed in a bottom-up approach containing a total number of 22 classes. The first level was associated with vegetation physiognomy (6 classes), the second level with linear features (6 classes) and the third level with land uses existing in the area (10 classes). Image objects were created with multiresolution segmentation, an algorithm supplied by eCognition software. The segmentation parameters were selected through a trial-and-error approach after visual evaluation of the resulting image objects. The rule-set comprised 100 classification rules described with the ‘Membership Function’ classifier. The classification stability was found to lie between 0.59 and 0.77, inversely proportional to the complexity of each level's classification. For an accuracy assessment, the error matrix method was used in a set of 250 randomly selected points. The overall classification accuracy achieved at the first level was 74%, at the second level 50% and at the third level 64%. The geometric accuracy of the classification was beyond the scope of this research; and moreover, consistent reference data sets were not available. The conclusion is that the use of rules in an object-based image analysis (OBIA) process has the potential to produce accurate landscape maps even in the case of complex environments, in which ancillary data are not available. Future work should focus on testing the transferability of the rule-set in different Mediterranean study sites, in order to draw a conclusion in relation to its potential operational use.     

Decision tree regression for soft classification of remote sensing data

In recent years, decision tree classifiers have been successfully used for land cover classification from remote sensing data. Their implementation as a per-pixel based classifier to produce hard or crisp classification has been reported in the literature. Remote sensing images, particularly at coarse spatial resolutions, are contaminated with mixed pixels that contain more than one class on the ground. The per-pixel approach may result in erroneous classification of images dominated by mixed pixels. Therefore, soft classification approaches that decompose the pixel into its class constituents in the form of class proportions have been advocated. In this paper, we employ a decision tree regression approach to determine class proportions within a pixel so as to produce soft classification from remote sensing data. Classification accuracy achieved by decision tree regression is compared with those achieved by the most widely used maximum likelihood classifier, implemented in the soft mode, and a supervised version of the fuzzy c-means classifier. Root Mean Square Error (RMSE) and fuzzy error matrix based measures have been used for accuracy assessment of soft classification.     

Land cover post-classifications by Markov chain geostatistical cosimulation based on pre-classifications by different conventional classifiers

The recently proposed Bayesian Markov chain random field (MCRF) cosimulation approach, as a new non-linear geostatistical cosimulation method, for land cover classification improvement (i.e. post-classification) may significantly increase classification accuracy by taking advantage of expert-interpreted data and pre-classified image data. The objective of this study is to explore the performance of the MCRF post-classification method based on pre-classification results from different conventional classifiers on a complex landscape. Five conventional classifiers, including maximum likelihood (ML), neural network (NN), Support Vector Machine (SVM), minimum distance (MD), and k-means (KM), were used to conduct land cover pre-classifications of a remotely sensed image with a 90,000 ha area and complex landscape. A sample dataset (0.32% of total pixels) was first interpreted based on expert knowledge from the image and other related data sources, and then MCRF cosimulations were performed conditionally on the expert-interpreted sample dataset and the five pre-classified image datasets, respectively. Finally, MCRF post-classification maps were compared with corresponding pre-classification maps. Results showed that the MCRF method achieved obvious accuracy improvements (ranging from 4.6% to 16.8%) in post-classifications compared to the pre-classification results from different pre-classifiers. This study indicates that the MCRF post-classification method is capable of improving land cover classification accuracy over different conventional classifiers by making use of multiple data sources (expert-interpreted data and pre-classified data) and spatial correlation information, even if the study area is relatively large and has a complex landscape.     

Object-based urban detailed land cover classification with high spatial resolution IKONOS imagery

Improvement in remote sensing techniques in spatial/spectral resolution strengthens their applicability for urban environmental study. Unfortunately, high spatial resolution imagery also increases internal variability in land cover units and can cause a ‘salt-and-pepper’ effect, resulting in decreased accuracy using pixel-based classification results. Region-based classification techniques, using an image object (IO) rather than a pixel as a classification unit, appear to hold promise as a method for overcoming this problem. Using IKONOS high spatial resolution imagery, we examined whether the IO technique could significantly improve classification accuracy compared to the pixel-based method when applied to urban land cover mapping in Tampa Bay, FL, USA. We further compared the performance of an artificial neural network (ANN) and a minimum distance classifier (MDC) in urban detailed land cover classification and evaluated whether the classification accuracy was affected by the number of extracted IO features. Our analysis methods included IKONOS image data calibration, data fusion with the pansharpening (PS) process, Hue–Intensity–Saturation (HIS) transferred indices and textural feature extraction, and feature selection using a stepwise discriminant analysis (SDA). The classification results were evaluated with visually interpreted data from high-resolution (0.3 m) digital aerial photographs. Our results indicate a statistically significant difference in classification accuracy between pixel- and object-based techniques; ANN outperforms MDC as an object-based classifier; and the use of more features (27 vs. 9 features) increases the IO classification accuracy, although the increase is statistically significant for the MDC but not for the ANN.     

Improving classical contextual classifications

This paper shows some combinations of classifiers that achieve high accuracy classifications. Traditionally the maximum likelihood classification is used as an initial classification for a contextual classifier. We show that by using different non-parametric spectral classifiers to obtain the initial classification, we can significatively improve the accuracy of the classification with a reasonable computational cost. In this work we propose the use of different spectral classifications as initial maps for a contextual classifier (ICM) in order to obtain some interesting combinations of spectral-contextual classifiers for remote sensing image classification with an acceptable trade-off between the accuracy of the final classification and the computational effort required.     

Use of textural measurements to map invasive wetland plants in the Hudson River National Estuarine Research Reserve with IKONOS satellite imagery

At this point, models, and accompanying field data, that could be used to predict the likely response of estuaries and tidal marshes to future environmental change are lacking. To improve this situation, monitoring efforts in these complex ecosystems need to be intensified, and new, efficient monitoring techniques should be developed. In this context, our research assessed the use of IKONOS satellite imagery to map plant communities at Tivoli Bays, in the Hudson River National Estuarine Research Reserve (HRNERR). Tivoli Bays, a freshwater tidal wetland, contains a unique assemblage of plant communities, including three invasive plants (Trapa natans, Phragmites australis, and Lythrum salicaria). To study the effects of textural information on the accuracy of land cover maps produced for the HRNERR, seven different 11-class land cover maps were produced using a maximum-likelihood classification on seven combinations of spectral and textural data derived from an IKONOS image. Conventional contingency tables served as a basis for an accuracy assessment of these maps. The overall classification accuracies, as assessed by the contingency tables, ranged from 45% to 77.7%. The maximum-likelihood classification relying on four spectral and four 5-by-5 filter textural bands (created by superposing a textural filter separately on each band of the IKONOS image) had the lowest overall accuracy, whereas the one based on four spectral and four 3-by-3 filter textural bands associated with all segments, identified by an object-based classification of the IKONOS image, had the highest accuracy. Results suggest that a combination of per-pixel classification and incorporation of texture for segments generated through an object-based classification slightly increases classification accuracy from 76.2% for the maximum-likelihood classification of the four spectral bands of the IKONOS image to 77.7% for the combination of spectral and textural information produced for selected segments. Further analysis indicates that better results may be obtained by using other types of data within the segments and that the traditional approach to the selection of training and accuracy sites may negatively bias the results for a combination per-pixel and object-based classification.     

ELM-based spectral–spatial classification of hyperspectral images using extended morphological profiles and composite feature mappings

Extreme Learning Machine (ELM) is a supervised learning technique for a class of feedforward neural networks with random weights that has recently been used with success for the classification of hyperspectral images. In this work, we show that the morphological techniques can be integrated in this kind of classifiers using several composite feature mappings which are proposed for ELM. In particular, we present a spectral–spatial ELM-based classifier for hyperspectral remote-sensing images that integrates the information provided by extended morphological profiles. The proposed spectral–spatial classifier allows different weights for both spatial and spectral features, outperforming other ELM-based classifiers in terms of accuracy for land-cover applications. The accuracy classification results are also better than those obtained by equivalent spectral–spatial Support-Vector-Machine-based classifiers.     

Application of spectral mixture analysis to Amazonian land-use and land-cover classification

Abundant vegetation species and associated complex forest stand structures in moist tropical regions often create difficulties in accurately classifying land-use and land-cover (LULC) features. This paper examines the value of spectral mixture analysis (SMA) using Landsat Thematic Mapper (TM) data for improving LULC classification accuracy in a moist tropical area in Rondônia, Brazil. Different routines, such as constrained and unconstrained least-squares solutions, different numbers of endmembers, and minimum noise fraction transformation, were examined while implementing the SMA approach. A maximum likelihood classifier was also used to classify fraction images into seven LULC classes: mature forest, intermediate secondary succession, initial secondary succession, pasture, agricultural land, water, and bare land. The results of this study indicate that reducing correlation between image bands and using four endmembers improve classification accuracy. The overall classification accuracy was 86.6% for the seven LULC classes using the best SMA processing routine, which represents very good results for such a complex environment. The overall classification accuracy using a maximum likelihood approach was 81.4%. Another finding is that use of constrained or unconstrained solutions for unmixing the atmospherically corrected or raw Landsat TM images does not have significant influence on LULC classification performances when image endmembers are used in a SMA approach.