Advances in classification for land cover mapping using SPOT HRV imagery

Abstract

High-resolution data from the HRV (High Resolution Visible) sensors onboard the SPOT-1 satellite have been utilized for mapping semi-natural and agricultural land cover using automated digital image classification algorithms. Two methods for improving classification performance are discussed. The first technique involves the use of digital terrain information to reduce the effects of topography on spectral information while the second technique involves the classification of land-cover types using training data derived from spectral feature space. Test areas in Snowdonia and the Somerset Levels were used to evaluate the methodology and promising results were achieved. However, the low classification accuracies obtained suggest that spectral classification alone is not a suitable tool to use in the mapping of semi-natural cover types.     

Noise removal for SPOT HRV imagery

Abstract

SPOT HRV imagery acquired in ‘double’ mode displays near vertical striping occurring every seven or eight pixels in the down-track scan direction. Six digital image restoration procedures have been examined for suppressing or removing this noise. Image restoration by filtering in the spatial frequency domain was found to be the most effective procedure. However, although less effective at noise removal, a simple median filtering procedure could be used with greatly reduced computational cost.     

Classification of SPOT HRV imagery and texture features

Abstract

Spatial co-occurrence matrices were computed for a SPOT HRV multispectral image for a moderate-relief environment in eastern Canada. The texture features entropy and inverse difference moment were used with the spectral data in landcover classification, and substantive increases in accuracy were noted. These range from 10 per cent for exposed bedrock to over 40 per cent in forest and wetland classes. The average classification accuracies were increased from 511 per cent (spectral data alone) to 86.7 per cent (spectral data plus entropy measured in band 2 and inverse difference moment in band 3). Classes that are homogeneous on the ground were characterized adequately by spectral tone alone, but classes containing mixed vegetation patterns or strongly related to structure were characterized more accurately by using a mixture of spectral tone and texture.     

Land-use classification of SPOT HRV data using a cover-frequency method

Abstract

A two-stage classification procedure has been applied to extract land use in a rural-urban fringe environment from SPOT High Resolution Visible (HRV) multi-spectral data. In this procedure, the SPOT HRV data were first classified into twelve land-cover types using a supervised maximum-likelihood classification (MLC). In the second stage, cover frequencies were extracted by moving a pixel window over the land-cover map obtained at the first stage. These cover frequencies were then employed in the classification of 14 land-use classes using a supervised minimum-city-block classifier. Results obtained with the cover-frequency method have been compared with those obtained using the conventional MLC approach. The overall accuracy measured by the Kappa coefficient was 0·462 for the MLC method; it was significantly improved to 0·663 with the cover-frequency method.     

Street orientation detection and recognition in Landsat TM and SPOT HRV imagery

Building (street) orientation is one of the important parameters for estimation of building bulk size (height and width) from corner reflector effects using remotely sensed radar image data. However, this parameter is difficult to obtain directly from radar data. Other sensor data such as optical and near infrared data may provide possibilities. This paper reports on a method for detection and recognition of street orientation in remotely sensed Landsat TM and/or SPOT HRV imagery. The methodology includes two steps: (1) multiscale wavelet transform techniques are employed to detect edges; (2) the predominant street orientation for each 20 × 20 pixel block is then recognised by applying a simple algorithm to the detected edges which contain most of the information about street orientations.     

Noise removal for SPOT HRV imagery; comments on the letter by Quarmby 1987

Abstract

The origins of a fault described previously in SPOT HRV imagery are explained. This fault caused serious vertical striping and steps have now been taken to ensure that data gathered by SPOT after 16 November 1986 are free of this.     

Field-based crop classification using SPOT4, SPOT5, IKONOS and QuickBird imagery for agricultural areas: a comparison study

A comparison of agricultural crop maps from independent field-based classifications of the Satellite Pour l'Observation de la Terre (SPOT) 4 multispectral (XS), SPOT5 XS, IKONOS XS, QuickBird XS and QuickBird pan-sharpened (PS) images is presented. An agricultural area within the north-west section of Turkey was analysed for field-based crop identification. The SPOT4 XS, SPOT5 XS, IKONOS XS and QuickBird images were collected in similar climatic conditions during July and August 2004. The classification of each image was carried out separately on a per-field basis on all bands and the coincident bands that are green, red and near-infrared (NIR). To examine the effect of filtering on field-based classification, the images were each filtered using the 3?×?3, 5?×?5, 7?×?7 and 9?×?9 mean filter and the filtered bands were also classified on per-field basis. For the unfiltered images, IKONOS XS provided the highest overall accuracies of 88.9% and 88.1% for the all-bands and the coincident bands classifications, respectively. On average, IKONOS XS performed slightly better than QuickBird XS and QuickBird PS, while it outperformed SPOT4 XS and SPOT5 XS. The use of filtered images in field-based classification reduced the accuracies for SPOT4 XS, SPOT5 XS, IKONOS XS and QuickBird XS. The results of this study indicate that smoothing images prior to classification does not improve the accuracies for the field-based classification. On the contrary, the accuracies for the filtered QuickBird PS images indicated a slight improvement. On the whole, both IKONOS and QuickBird images produced quite promising results for field-based crop mapping, yielding overall accuracies above 83%.     

Improvement of forest type classification by SPOT HRV with 20 m mesh DTM

Abstract

For multispectral analysis of forest land in mountainous areas, the estimation of true reflectance without the terrain having an effect on the sensor response is indispensable. To study this subject, the authors carried out the following experiment. First, we made a precise digital terrain model (DTM) at an interval of 10 m for a test forest site covered with Lambertian-type crown surface. Analysing the forest land from the SPOT data with the precise DTM, we obtained a classification result of forest type about 20 per cent higher accuracy than the result without application of this method.     

Integrating automatically processed SPOT HRV Pan imagery in a DEM-based procedure for channel network extraction

The contribution of automatically processed Système Pour l'Observation de la Terre (SPOT) high resolution visible (HRV) Pan data as an ancillary source of information to a digital elevation model (DEM)-based method for channel network extraction is introduced. The image processing stage included an application of a Laplacian filter for edge detection. Edge pixels that were not contiguous with the main channel network were eliminated and the channel network was buffered and finally skeletonized to create channels with one-pixel width. A DEM-based approach was implemented for an overlapping area using the terrain analysis using digital elevation models (TauDEM) procedure based on the principles of the flow direction matrix method. The channel network that was extracted by the use of the two methods in fusion was tested against the imagery, and the DEM-based, channel networks and conformed to the reference data more accurately in terms of coverage of channels, network connectivity and location of extracted channels. A disadvantage of the data fusion is the additional, few, artificial channels.     

Preliminary analysis of SPOT HRV multispectral products of an arid environment

A preliminary analysis of a SPOT HRV multispectral scene centred on the Chott el Djerid in southern Tunisia is presented. All three HRV bands are very strongly correlated for this scene, and statistically it has an overall one-dimensional structure, although the near-infrared band provides unique information on the variability within vegetated areas. Each of the three bands has discriminatory potential. Inter-detector variability is clearly visible in HRV band 2, and there are also systematic changes in level. The improved spatial resolution of the HRV sensor is notable, compared with the Thematic Mapper.     

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.     

Neural classification of SPOT imagery through integration of intensity and fractal information

It is well known that higher dimensional information essentially leads to better accuracy in remotely sensed image classification. This paper is aimed at land cover classification from SPOT-HRV imagery by the integration of multispectral intensity and texture information. In particular, fractal dimensions are extracted using a wavelet transform as image texture. A neural network approach to classification is adopted in this paper. The underlying network is a modified multilayer perceptron trained by a Kalman filtering technique. The main advantages of this network are (1) its non-backpropagation fashion of learning which leads to a fast convergence, (2) a built-in optimization function, and (3) global scale. Saving computer storage space and a fast learning capability are in particular suitable features for remote sensing applications. Correlation analysis was subsequently performed on both the intensity and fractal images. It was found that fractal information significantly improves the discrimination capability of heterogeneous area such as in urban regions, while it slightly degrades accuracy for homogeneous areas, such as open water. The overall classification performance is superior to results obtained using reflectance only. Improvements over heterogeneous areas are demonstrated.     

Remote-sensing imagery classification using multiple classification algorithm-based AdaBoost

AdaBoost demonstrates excellent performance in remote sensing (RS) image classification, but as it works on only one classification algorithm, the disadvantage of the classification algorithm itself is difficult to overcome, resulting in limitations in the improvement of classification accuracy. In this article, a modified AdaBoost, a multiple classification algorithm-based AdaBoost (MCA AdaBoost), is proposed to improve remote sensing image classification. The new method works on more than one classification algorithm and can make full use of the advantages of different learning algorithms. Based on a Landsat 8 Operational Land Imager (OLI) image whose spatial resolution was enhanced to 15 m with a panchromatic band, a C4.5 decision tree, Naïve Bayes, and artificial neural network were used as objects to verify and compare the performance of both AdaBoost and MCA AdaBoost. The experimental results show that MCA AdaBoost successfully inherits the benefits of the original AdaBoost, combines the advantages of different classification algorithms and lowers overfitting. By increasing diversity and complementarity among base classifiers, MCA AdaBoost outperforms AdaBoost in terms of RS classification accuracy improvement.     

Filtering SPOT imagery by kriging analysis

Digital images are rich in data, but in many instances they are so complex as to require spatial filtering to distinguish the structures in them and facilitate interpretion. The filtering can be done geostatistically by kriging analysis. It proceeds in two stages. The first involves modelling the correlation structure in the imagery by decomposing the variogram into independent spatial components. The second takes each component in turn and kriges it, thereby filtering it from the others. The paper describes the theory and illustrates it with an example of an analysis of a SPOT image in a forested landscape of the south-eastern United States. Variograms of the three wavebands, originally recorded as digital numbers and for the red and infrared transformed to the logarithms, revealed spatial variation on two distinct scales with effective ranges of 300m and 3km. These variograms and that of the Normalized Difference Vegetation Index (NDVI) were fitted by nested (double) exponential models. The two spatial components in the scene were then estimated separately by kriging analysis and mapped. The maps of NDVI are displayed and compared with data from ground survey. The shortrange component represents an intricate pattern of dissection and its associated vegetation. The long-range component is that of the major landform units and associated ground cover classes.     

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.     

Multi-spectral classification of snow using NOAA AVHRR imagery

Abstract

Problems of accurate discrimination between snow and cloud, together with the detection of the snow pack boundary, have handicapped the use of satellite data in operational snow-cover mapping systems. A technique, involving an unsupervised clustering procedure, is described which allows the removal of cloud areas using NOAA-9 Advanced Very High Resolution Radiometer (AVHRR) channel-1, channel-3 and channel-4 data in conditions of recent snow lie and a difference channel (channel-2 —channel-1 with channel-3 and channel-4) during periods of advanced snow melt. Accurate delineation of snow extent is provided by the techniques if these specified snow conditions are taken into account. A method for the identification of areas of marginal snow melt is also presented, based on comparisons with Landsat Thematic Mapper data. The classifications also enable the determination of snow areas influenced by cloud shadows and conifer forest in addition to separating areas of differing snow depth and percentage cover.     

Supporting collaborative hierarchical classification: Bookmarks as an example

Bookmarks (or favorites, hotlists) are popular strategies to relocate interesting websites on the WWW by creating a personalized URL repository. Most current browsers offer a facility to locally store and manage bookmarks in a hierarchy of folders; though, with growing size, users reportedly have trouble to create and maintain a stable organization structure. This paper presents a novel collaborative approach to ease bookmark management, especially the “classification” of new bookmarks into a folder. We propose a methodology to realize the collaborative classification idea of considering how similar users have classified a bookmark. A combination of nearest-neighbor-classifiers is used to derive a recommendation from similar users on where to store a new bookmark. A prototype system called CariBo has been implemented as a plugin for the central bookmark server software SiteBar. All findings have been evaluated on a reasonably large scale, real user dataset with promising results, and possible implications for shared and social bookmarking systems are discussed.     

Mapping coffee plantations with Landsat imagery: an example from El Salvador

Considering the potential of shaded coffee plantations mixed with natural vegetation for promoting biodiversity conservation, this project assessed the utility of multi-date Landsat Thematic Mapper (TM) satellite imagery for the characterization of natural vegetation versus coffee plantations in western El Salvador. For assembling a multi-temporal Landsat TM data set, we applied a regression analysis model to remove cloud cover and cloud shadows. Then, through a hybrid classification approach, a nine-class land use/land cover (LULC) map was generated. We identified two types of coffee plantations (‘open-canopy’ and ‘close-canopy’) along with natural forest/shrubland, mangrove, water bodies, sandy coastal soils, bare soil, urban areas and agriculture. Notwithstanding the small sample size of the accuracy data, our assessment revealed an overall accuracy of 76.7% (Kappa coefficient?=?0.68), considering only the four classes with independent field data. The overall classification accuracy for distinguishing coffee plantations from non-mangrove natural forest was 81.6% and the classification accuracy for distinguishing ‘open-canopy’ from ‘close-canopy’ coffee plantations was 85.7%. We are encouraged by the results of this prototype study. They indicate that remote-sensing techniques can be used to distinguish different classes of coffee production systems and to differentiate coffee from natural forest.