Comparaison de differentes methodes de traitement par analyses en composantes principales des donnees Thematic Mapper de LANDSAT-5 appliquees al'inventaire forets mediterranneennes du sud de la France

Abstract

The existence of remotely sensed data with high spatial resolution, like the ones produced by the Thematic Mapper of LANDSAT, set the problem of the reduction of the number of spectral dimensions to be analysed. The work presented hereby is in this context and aims to compare the performances of the supervized classification applied to the perception of forested and sub-forested mediterranean ecosystems processed with: original TM data selected only under spectral consideration; components factors provided by a classical Principal Component Analysis (PCA); components factors provided by a selective Principal Component Analysis (PCA). The analysis of the results shows that the classification done using the three axes extracted by a classical Principal Component Analysis of the six Thematic Mapper bands gives better results than all the other combinations (original data or data provided by selective PCA). On the other hand, and for all the classifications processed, it appears that the performances are excellent (near 90 per cent) for units representing stable systems, climatic or definitely degraded. At the opposite end, performances are quite good (near 60 per cent) for the evolutive systems (in regression or progression).     

Performance of LANDSAT-5 TM data in land-cover classification†

An experimental analysis has been conducted on the performance of the new LANDSAT-5 Thematic Mapper (TM) data for detailed land-cover classification using a maximum-likelihood method. Data used is the TM test data of the Tokyo metropolitan area (path-107, row-035) of 4 November, 1984. Map-precision geometric correction is performed and TM data are resampled to 30 m pixel spacing. The experiment is designed to determine how well TM categories land-cover types in comparison with the Detailed Numerical Information digitally formatted data (Geographical Survey Institute of Japan, 10 m spatial accuracy), together with ground truth data in a representative test area. Classification accuracy for aggregated 12 categories within the test area is about 47 per cent with the application of the explicit filtering technique utilizing 3×3 neighbourhood operations. This increases to 70 per cent using a majority logic filter with a larger 5×5 neighbourhood function. Associated with the classification accuracy, effects of the mixed pixels are also investigated. The results show that the improved characteristics of TM aided the overall classification accuracy.     

基于模糊函数多域特征融合与集成学习的雷达辐射源信号识别

针对复杂电磁环境下雷达辐射源信号识别方法中存在的抗噪性能差、识别准确率低等问题,提出一种融合模糊函数多域投影特征的集成深度学习识别方法.首先,对信号的模糊函数进行高斯平滑处理,从多域视角出发选取合适角度对模糊函数进行二维投影以构建特征数据集;然后,构建一种基于多域特征融合的两阶段识别分类方法,使用多个密集连接网络DenseNet 121作为初级分类器分别对3类特征数据集进行训练学习,得到初级分类结果;最后,通过Stacking策略对初级分类结果进行融合学习,得到最终类别信息.实验结果表明,所提出方法在信噪比为0 dB时对6类典型雷达信号的整体平均识别率均保持在97.24%以上,即使是在-4 dB环境中,识别率也稳定在87.16%以上,验证了所提出方法的有效性和可行性,具有一定的工程价值.     

单脉冲雷达合并通道定向接收机建模

对一种在单脉冲雷达中采用的以调相信号的频率压缩和合并通道中压缩信号的限幅为基础的通道合并定向接收机(角度鉴别器)进行了分析，建立了这种定向接收机在三通道幅相特性不平衡时的仿真用模型，应用这一模型可以分析接收机非线性、通道间耦合、和差通道幅相特性的不平衡等对单脉冲雷达定向特性的影响。     

Radar backscattering of forest stands†

Research on the potential applications of microwave remote sensing in agriculture is conducted in the Netherlands by the ROVE team. Since active microwave remote sensing, featuring its all-weather capability, also seems to be a promising tool for forest classification, especially on a global scale, the Wageningen Agricultural University started a new working group in co-operation with the ROVE team in order to explore this field of application.

Results of four X-band SLAR flights have been analysed. The digital radar images obtained are accurately corrected both geometrically and radiometrically and indicate gamma values instead of arbitrary grey tones.

Radar signatures, showing seasonal and angular effects, of 16 classes of forest stands have been derived from the images. Special attention has been paid to the statistical properties of the radar signatures and their impact on classification accuracy. Several interesting phenomena have been observed indicating effects of vegetation structure on radar backscattering.

One of the test areas is a young forest in the Oost-Flevoland polder featuring a substantial variety of species; parcels are relatively large, rectangular in shape, homogeneous in structure and age and with pure species stands making this an ideal test site. Another test area is an old forest located at the Veluwe. Much variation in age is present here, which made it possible to determine relationships between tree age and radar backscatter for several coniferous tree species.

The initial work as presented here clearly demonstrates the appropriateness of X-band ability in the classification of (Dutch) forests. Theoretical considerations suggest that a multitemporal approach is likely to give the most accurate results of tree-type classification. A classification simulation yielded overall error fractions ranging from 10 to 16 per cent at the Oost-Flevoland polder test area and 14 to 28 per cent at the Veluwe test area. This can be demonstrated in multitemporal radar images as well as in actual classified images.     

Spatial discrimination of salt- and sodium-affected soil surfaces

Salinization-alkalinization is a time- and space-dynamic soil degradation process in semiarid regions. This study implements a synergistic approach to map salt- and sodium affected surfaces, combining digital image classification with field observation of soil degradation features and laboratory determinations. Salinity-alkalinity classes were established using the electrical conductivity (EC) and pH values. A neighbourhood operator, with spatial and spectral user-defined constraints determined the spectral objects constituting the training set. Six combined Landsat TM bands (1,2,4,5,6,7) provided the highest separability between salt- and sodium-affected soil classes. Although the overall accuracy was slightly low (64 per cent), accuracies of 100 per cent were obtained for some classes. Main causes of spectral confusions, masking different salinity-alkalinity degrees were the type and abundance of salt-tolerant vegetation cover, the topsoil textures, and the mixture of topsoil properties under field conditions.     

Object-based urban structure type pattern recognition from Landsat TM with a Support Vector Machine

This study evaluates the potential of object-based image analysis in combination with supervised machine learning to identify urban structure type patterns from Landsat Thematic Mapper (TM) images. The main aim is to assess the influence of several critical choices commonly made during the training stage of a learning machine on the classification performance and to give recommendations for classifier-dependent intelligent training. Particular emphasis is given to assess the influence of size and class distribution of the training data, the approach of training data sampling (user-guided or random) and the type of training samples (squares or segments) on the classification performance of a Support Vector Machine (SVM). Different feature selection algorithms are compared and segmentation and classifier parameters are dynamically tuned for the specific image scene, classification task, and training data. The performance of the classifier is measured against a set of reference data sets from manual image interpretation and furthermore compared on the basis of landscape metrics to a very high resolution reference classification derived from light detection and ranging (lidar) measurements. The study highlights the importance of a careful design of the training stage and dynamically tuned classifier parameters, especially when dealing with noisy data and small training data sets. For the given experimental set-up, the study concludes that given optimized feature space and classifier parameters, training an SVM with segment-shaped samples that were sampled in a guided manner and are balanced between the classes provided the best classification results. If square-shaped samples are used, a random sampling provided better results than a guided selection. Equally balanced sample distributions outperformed unbalanced training sets.     

An evaluation of Landsat TM spectral data and SAR-derived textural information for lithological discrimination in the Red Sea Hills,Sudan

The effectiveness of spectral and textural information in the identification of surface rock types in an arid region, the Red Sea Hills of Sudan, is evaluated using spectral information from the six Landsat TM optical bands and textural features derived from Shuttle Imaging Radar-C (SIR-C) C-band HH polarization data. An initial classification is derived from Landsat TM data alone using three classification algorithms, Gaussian maximum likelihood, a multi-layer feed-forward neural network and a Kohonen self-organizing feature map (SOM), to generate lithological maps, with classification accuracy being measured using a confusion matrix approach. The feed-forward neural net produced the highest overall classification accuracy of 57 per cent and was, therefore, selected for the second experiment, in which texture measures from SIR-C C-band HH-polarized synthetic aperture radar (SAR) data are added to selected TM spectral features. Four methods of measuring texture are employed, based on the Fourier power spectrum, grey level co-occurrence matrix (GLCM), multi-fractal measures, and the multiplicative autoregressive random field (MAR) model. The use of textural information together with a subset of the TM spectral features leads to an increase in classification accuracy to almost 70 per cent. Both the MAR model and the GLCM matrix approach perform better than Fourier and multi-fractal based methods of texture characterization.     

NOAA-AVHRR NDVI decomposition and subpixel classification using linear mixing in the Argentinean Pampa

Crop Normalized Difference Vegetation Index (NDVI) time profiles and crop acreage estimates were derived from the application of linear mixture modelling to Advanced Very High Resolution Radiometer (AVHRR) data over a test area in the southern part of the Pampa region, Argentina. Bands 1 and 2 from seven AVHRR scenes (June to January 1991) were combined to produce fraction images of winter crops, summer crops and pastures. A Landsat Thematic Mapper (TM) scene of the region was classified and superimposed to the AVHRR Local Area Coverage (LAC) data by means of a correlation technique. Each class signature was extracted by regressing the AVHRR response on the cover types proportions, estimated from Landsat-TM data, over sets of calibration windows. The crop NDVI profiles were hence derived from the class signatures in bands 1 and 2. These profiles appeared consistent with the cover types, but variability depending on the set of windows was noted. The assessment of the class signatures was indirectly accomplished through the subpixel classifications of the AVHRR data, performed using the different sets of class spectra. Although some discrepancies between AVHRR and Landsat–TM estimates were observed at the individual window level, the classification results compared quite well on a regional scale with Landsat–TM estimates: crop acreage was estimated to an overall accuracy ranging from 89 to 95 per cent according to the spectra used in the classification. Definitely, the proposed methodology should permit a better exploitation of the temporal resolution of AVHRR data in both the areas of yield prediction and vegetation classification. Furthermore, the perational application of such a methodology for crop monitoring will undoubtedlybe facilitated with the coming sensor systems such as the ModerateResolution Imaging Spectroradiometer (MODIS), the SPOT Vegetation Monitoring Instrument or the ‘Satelite Argentino Cientifico’ (SAC–C).     

Shuttle radar mapping with diverse incidence angles in the rainforest of Borneo

Abstract

Shuttle Imaging Radar-B (SIR-B) images, obtained at two different incidence angles were analysed for discrimination and mapping of vegetation in the rainforest of Borneo. In an area of coastal lowland three units of forest canopy and two units of open surface cover were distinguished and mapped. The backscatter characteristics of the units were compared and analysed by ratioing digital image data of sample sites against digital data of reference sites. The ratios show that the backscatter for each surface unit is clustered in separate, mostly non-overlapping, domains at both incidence angles. Different rates of change with incidence angle indicate a strong angular dependence in the dominant backscatter mechanism of swamp that is not apparent for the other units. A corresponding digitally coregistered Landsat multispectral scanner (MSS) image helped to verify the spatial distribution of the mapped units in the coastal lowland. In the interior upland three major forest species associations that have contrasted canopy structures were not discriminated on the SIR-B images. Owing to perennial cloud cover the distribution and extent of the different canopies cannot be determined from Landsat MSS or other images obtained at optical wavelengths. The ability to discriminate and map different forest canopies in the interior of Borneo requires a wider radar response capability which may be obtainable at shorter wavelengths and with multipolariz-ation states.     

The effects of spatial resolution on the classification of Thematic Mapper data

Actual and degraded LANDSAT-4 Thematic Mapper (TM) data were analysed to examine the effect of spatial resolution on the performance of a per pixel, maximum-likelihood classification algorithm. Analysis of variance (ANOVA) and a balanced, three-factor, eight-treatment, fixed-effects model were used to investigate the interactions between spatial resolution and two other TM refinements, spectral band configuration and data quantization. The goal was to evaluate quantitatively the effects of these attributes on classification accuracies obtained with all pixels (pure pixels plus mixed pixels) and on accuracies obtained with pure pixels alone.

A comparison of results from these separate analyses supported previous explanations of the effects of increasing spatial resolution. First, the difficulty in classifying mixed pixels was demonstrated by an average 21 per cent decrease in percentage accuracy from the pure-pixel case to the pure-plus-mixed-pixel case for the eight ANOVA treatments. In the pure-pixel case, an increase in spatial resolution from 80 to 30 m caused an average 6·1 per cent decrease in percentage accuracy when the other factors were held constant. This decrease was attributed to increased within-class spectral variability at the TM resolution. Finally, in the pure-plus-mixed-pixel case, increasing the spatial resolution did not significantly affect accuracy. This insignificance was attributed to a reduced proportion of mixed pixels at the TM resolution which counteracted the detrimental effects of increased spectral variability. These results point to a need for the development of new approaches to classification which take full advantage of the TM spatial resolution.     

Simulated and observed L-HH radar backscatter from tropical mangrove forests

Abstract

We applied the Santa Barbara canopy backscatter model to model radar backscatter from mangrove forest stands in the Ganges delta of southern Bangladesh, and assessed the feasibility of delineating flooding boundaries within the stands. Modelled L-band (0-235 m wavelength) HH backscatter showed that canopy volume scattering dominated for stands under nonflooded ground surface. Double bounce trunk-ground term were enhanced by the presence of water under trees. For flooded mangrove forest, the trunk-ground term was dominant at small radar incidence angles; the trunk-ground term dominancy reduced as the incidence angle increased. Shuttle Imaging Radar (SIR-B) data and model results showed that for the mangrove forest, radar data with small incidence angles should be used to delineate the flooding boundaries.     

SAR image processing using adaptive stack filter

Stack filters are a special case of non-linear filters. They have a good performance for filtering images with different types of noise while preserving edges and details. A stack filter decomposes an input image into several binary images according to a set of thresholds. Each binary image is filtered by a Boolean function. The Boolean function that characterizes an adaptive stack filter is optimal and is computed from a pair of images consisting of an ideal noiseless image and its noisy version. In this work the behavior of adaptive stack filters on synthetic aperture radar (SAR) data is evaluated. With this aim, the equivalent number of looks for stack filtered data are calculated to assess the speckle noise reduction capability of this filter. Then a classification of simulated and real SAR images is carried out on data filtered with a stack filter trained with selected samples. The results of a maximum likelihood classification of these data are evaluated and compared with the results of classifying images previously filtered using the Lee and the Frost filters.     

Analysis of digital LANDSAT MSS and SEASAT SAR data for use in discriminating land cover at the urban fringe of Denver,Colorado

Digital SEASAT synthetic aperture radar (SAR) and LANDSAT multispectral scanner (MSS) data were evaluated to determine their utility to discriminate suburban and regional cover in the eastern fringe area of the Denver, Colorado, metropolitan area. The primary emphasis of the study was land-cover discrimination performance of MSS versus SAR and SAR/MSS combined. In addition, both a median-filtering and a data-smoothing procedure were tested in an attempt to increase the spectral separability between land-use/land-cover classes for SAR data. The results indicated that analysis of LANDSAT MSS data alone provided a significantly (α = 0·05) higher overall classification accuracy or improved spectral class separation than the best SEASAT SAR classification. However, when using LANDSAT MSS and SEASAT SAR data simultaneously, a significant increase in classification accuracy was obtained. Analysis indicated that SEASAT SAR data provided a measure of surface geometry that complemented the reflective characteristics of LANDSAT MSS visible and near-infrared data. Smoothing and median filtering provided significant improvement in classification accuracy over non-filtered SAR data.     

Automatic gender detection using on-line and off-line information

In this paper, the problem of classifying handwritten data with respect to gender is addressed. A classification method based on Gaussian Mixture Models is applied to distinguish between male and female handwriting. Two sets of features using on-line and off-line information have been used for the classification. Furthermore, we combined both feature sets and investigated several combination strategies. In our experiments, the on-line features produced a higher classification rate than the off-line features. However, the best results were obtained with the combination. The final gender detection rate on the test set is 67.57%, which is significantly higher than the performance of the on-line and off-line system with about 64.25 and 55.39%, respectively. The combined system also shows an improved performance over human-based classification. To the best of the authors’ knowledge, the system presented in this paper is the first completely automatic gender detection system which works on on-line data. Furthermore, the combination of on-line and off-line features for gender detection is investigated for the first time in the literature.     

Influence of buried topography on surface drainage pattern

Abstract

An automated system has been developed for mosaicking spaceborne synthetic aperture radar (SAR) imagery. The system is capable of producing multiframe mosaics for large-scale mapping by combining images in both the along-track direction and adjacent cross-track swaths from ascending and descending passes. The system requires no operator interaction and is capable of achieving high registration accuracy. The output product is a geocoded mosaic on a standard map grid such as UTM or polar stereographic. The procedure described in detail in this paper consists essentially of remapping the individual image frames into these standard grids, frame-to-frame image registration and radiometric smoothing of the seams. These procedures are directly applicable to both the Magellan Venus Mapper and a scanning SAR design such as Radarsat, Eos SAR in addition to merging image frames from traditional SAR systems such as SEASAT and SIR-B. With minor modifications, it may also be applied to spaceborne optical sensor data to generate large-scale mosaics efficiently and with a high degree of accuracy. The system has been tested with SEASAT, SIR-B and Landsat TM data. Examples presented in this paper include a 38-frame mosaic of the Yukon River basin in central Alaska, a 33-frame mosaic of southern California and a three-frame terrain-corrected geocoded mosaic of the Wind River basin in Wyoming.     

Preliminary analyses of SIR-B radar data for recent Hawaii lava flows

The Shuttle Imaging Radar (SIR-B) experiment acquired two L-band (23 cm wavelength) radar images (at about 28° and 48° incidence angles) over the Kilauea Volcano area of southeastern Hawaii. Geologic analysis of these data indicates that, although as lava flows and pyroclastic deposits can be discriminated, pahoehoe lava flows are not readily distinguished from surrounding low return materials. Preliminary analysis of data extracted from isolated flows indicates that flow type (i.e., as or pahoehoe) and relative age can be determined from their basic statistics and illumination angle.     

Application of the semivariogram textural classifier (STC) for vegetation discrimination using SIR-B data of Borneo

Abstract

Classification of Shuttle Imaging Radar-B (SIR-B) data from a rainforest-covered portion of Borneo is performed using image texture. The algorithm used is the semivariogram textural classifier (STC). This is a deterministic, supervised parallelepiped type classifier which provides the option of combining textural and radiometric information. Textural information is expressed by the semivariogram function. Radiometric information is conveyed by the mean digital number (DN) value. Results of the classificaiion cmulale a previously published map obtained by visual interpretation of the same SIR-B data set.     

考虑不平衡指数的不平衡数据集分类设计方法

数据集类别不平衡问题是分类领域的重要问题之一,每个数据集的不平衡指数都与其自身有着紧密的联系,是数据集的重要标志。面对不平衡数据集分类设计问题,提出了一种改进AdaBoost算法（enhanced AdaBoost,E-AdaBoost）。该算法将不平衡指数和不平衡数据集中较为重要的少数类分类正确率考虑到算法的迭代过程中,改进了基分类器的权重更新策略,进而提高对不平衡数据集的分类性能。基于E-AdaBoost的不平衡数据集分类设计方法可以根据样本的不平衡指数,确定基分类器的权重参数,进而提高分类器性能。利用该方法,结合多个经典分类器,在人工数据集和标准数据集上进行实验分析,并对比相关方法,结果表明,基于E-AdaBoost的不平衡数据集分类设计方法能够有效提高不平衡数据集的分类性能。     

基于粗糙集约简的决策林构建方法

针对如何提高决策林的分类精度问题,提出一种基于粗糙集约简构建决策林的技术,包括基于逐次数据约筒构建粗糙决策林和基于遗传算法构建粗糙决策林。对3个UCI数据集的验证表明,基于遗传算法构建的粗糙决策林获得了更好的分类效果。