Multispectral Classification Research Articles

多重分光画像分類のための最良線形判別関数の導出

多重分光画像分類のための最良線形判別関数の導出

Gaussian Maximum Likelihood and Contextual Classification Algorithms for Multicrop Classification

In this paper we review some of the ways in which context has been handled in the remote-sensing literature, and we introduce additional possibilities. The problem of computing exhaustive and normalized class-membership probabilities from the likelihoods provided by the Gaussian maximum likelihood classifier (to be used as initial probability estimates to start relaxation) is discussed. An efficient implementation of probabilistic relaxation is proposed, suiting the needs of actual remote-sensing applications. A modified fuzzy-relaxation algorithm using generalized operations between fuzzy sets is presented. Combined use of the two relaxation algorithms is proposed to exploit context in multispectral classification of remotely sensed data. Results on both one artificially created image and one MSS data set are reported.

A preliminary assessment of Landsat TM imagery for mapping vegetation and sediment distribution in the Wash estuary

Abstract Analysis of Landsat Thematic Mapper (TM) image of 14 May 1984 has shown that such data can be used to survey vegetation and sediment distributions in the intertidal zone of the Wash estuary at a spatial detail comparable with current methods practised by the.Nature Conservancy Council. Multispectral classification of this TM image showed good separation of salt-marsh vegetation communities which had recently been surveyed by the Nature Conservancy Council and for which reliable training data could be taken. The sensitivity of classification performance, using both parametric and non-parametric algorithms, to apparently minor differences in phenology at training site locations demonstrates two requirements for improved salt-marsh classification. They are the need for strict definition of training data and that TM wave bands 2, 3, 4 and 5 provide suitable spectral vectors for classifying intertidal environments.

Remote Sensing Leiters Elimination de centres initiaux inutiles pour l'algorithme aggregation autour de centres mobiles

Abstract He and Wang (1987) proposed the introduction of initial centres for the algorithm of aggregation around mobile centres. This involves the combination of peaks located in uni-dimensional histograms of data and provides not only the initial centres but also the empty centres to which no pixel will be located. When the number of empty centres is large, the lost calculating time becomes important. This letter reports a continuation of the previous study. Pre-processing is proposed to eliminate the empty centres before the first iteration of the algorithm of aggregation. To evaluate the performance of this pre-processing, we have realized a multispectral classification with and without the pre-processing on a sub-image of LANDSAT Thematic Mapper data. The result shows that this pre-processing improves the algorithm.

Noise in Remote-Sensing Systems: The Effect on Classification Error

Several types of noise in remote-sensing systems are treated. The purpose is to provide enhanced understanding of the relationship of noise sources to both analysis results and sensor design. The context of optical sensors and spectral pattern recognition analysis methods is used to enable tractability for quantitative results. First, the concept of multispectral classification is reviewed. Next, stochastic models are discussed for both signals and noise, including thermal, shot and quantization noise along with atmospheric effects. A model enabling the study of the combined effect of these sources is presented, and a system performance index is defined. Theoretical results showing the interrelated effects of the noise sources on system performance are given. Results of simulations using the system model are presented for several values of system parameters, using some noise parameters of the Thematic Mapper scanner as an illustration. Results show the relative importance of each of the noise sources on system performance, including how sensor noise interacts with atmospheric effects to degrade accuracy

Optimum Spatial Resolution for Multispectral Classification

Optimum Spatial Resolution for Multispectral Classification

Reduction of random noise from multiband image data using phase relationships among their Fourier coefficients

The problem of reducing the random noise from multiband image data is examined, taking a nine-band image data set of a terrestrial scene and adding a significant amount of noise to the original pixel value of each band. For each of these data sets, data of three nearby bands generally exhibit a high correlation among them. Selected data for several sets of three nearby bands are transformed from the spatial domain to the frequency domain to study phase relationships (i.e., coherency) among their Fourier coefficients of various frequencies. It is shown that, in general, the addition of random noise results in the rapid change, with frequency, of a quantity called the coherency measure. (This is a quantitative measure of the phase agreement among the phases of various Fourier coefficients at a given frequency.) The coherency measure vs. frequency curve for a given data line is then used to attenuate various Fourier coefficients of the corresponding nearby bands of that line. It is then shown that the inverse transformation of such modified Fourier coefficients results in a statistically significant reduction of noise from the data of single lines, or from those data of some finite areas of the image. Results of a supervised boxcar multispectral classification with the original as well as various modified data sets of the selected image are also presented to provide additional guidance in the use of such a sophisticated analytic procedure in image processing.

LANDSAT-4 MSS And Thematic Mapper Data Quality And Information Content Analysis

Landsat-4 Thematic Mapper and Multispectral Scanner data were analyzed to obtain information on data quality and information content. Geometric evaluations were performed to test band-to-band registration accuracy. Thematic Mapper overall system resolution was evaluated using scene objects which demonstrated sharp high contrast edge responses. Radiometric evaluation included detector relative calibration, effects of resampling, and coherent noise effects. Information content evaluation was carried out using clustering, principal components, transformed divergence separability measure, and numerous supervised classifiers on data from Iowa and Illinois. A detailed spectral class analysis (multispectral classification) was carried out on data from the Des Moines, Iowa area to compare the information content of the MSS and TM for a large number of scene classes.

Some experiments with spatial feature extraction methods in multispectral classification

Abstract Feature extraction is an important factor in determining the accuracy that can be attained in the classification of multispectral images. The traditional per point classification methods do not use all the available information, since they disregard the spatial relationships that exist among pixels belonging to the same class. In this paper, methods are developed to extract additional image spatial features by means of linear and non-linear local filtering. Feature selection methods are also developed, since it is usually not possible to use all the generated features. The classification stage is performed in a supervised mode using the maximum likelihood criterion. A quantitative analysis of the performance of the spatial features show that an overall increase in precision of classification is achieved, although at the expense of increased rejection levels, particularly on the borders between different fields.

Observation from landsat of ocean circulation

Landsat data were used to observe variations in ocean colour in the south-western Tasman Sea adjacent to the Australian coast and in Bass Strait. These variations were interpreted as ocean-circulation patterns and these patterns were seen to be consistent with the known behaviour of the water masses in the area. The satellite data were displayed for interpretation as grey-scale and pseudo-colour images of individual bands of the raw data, as false-colour composites of three bands of the raw data, as thematic maps following multispectral classification, and as grey-scale and pseudo-colour images of data corrected for the effects of atmospheric radiance. These images were interpreted visually, appealing to the known characteristics of the satellite sensor and the typical spectra of upwelling ocean radiances.

Des néo-variables polygéniques au secours d'une image hivernale

Polygenic neo-variables to the rescue of a winter scene. — This article seeks to demonstrate the potential application of remote sensing via a simulated mission of the satellite Spot for elucidating the urban landscape. The case is Strasbourg, the date, February 1982. Certain faults in this simulation exercice have unfortunately precluded the normal possibility of multi-spectral classification. Some attempts have been made to rectify these deficiencies by creating three new band categories: radiometric, structural, and textural. The morphological neo-variables of the two latter categories have provided indispensable help in producing a reasonably satisfactory classification. They have also raised a series of stimulating questions. There were other deficiencies besides those of the initial imagery which made it quite difficult to accomplish this particular case study. There was no specialized information which might facilitate corroborative analysis of the image. The fact that these unfavourable conditions have been surmounted should make for optimism about the potential use of better data from satellite Spot even in research centers which have only a modest level of graphic equipment.

Image processing for remote sensing

In this paper a number of approaches to multispectral image segmentation and classification are considered. The methods range from the simple Bayesian decision rule for classification of image data on pixel-by-pixel basis, to sophisticated algorithms using contextual information. Both the spatial pixel category dependencies and the two-dimensional correlation-type contextual information have been incorporated in decision-making schemes. The aim of these algorithms is to achieve a greater reliability in the process of interpretation of remote-sensing data.

A means for utilizing ancillary information in multispectral classification

A method is presented that allows information from ancillary data sources to be incorporated into the results of an existing classification of remotely sensed data. Based upon probabilistic label relaxation procedures, which are used for imbedding spatial context data in image-labeling problems, the method utilizes the source of ancillary information in the form of a set of probabilities. These are injected into a modified relaxation method called supervised relaxation labeling which, on application, develops a labeling for remotely sensed data that strikes a balance in consistency between spectral, spatial, and ancillary data sources of information. Results are presented of a forestry classification in which accuracy is improved from 68% to 81% by incorporating topographic elevation in the manner outlined.

Importance of Higher-Order Components to Multispectral Classification

A Landsat multispectral image was combined with the corresponding digital terrain elevation data to study several information extraction procedures. Principal component and limited multispectral classification procedures were conducted on 1024 × 1024 four-band Landsat and five-band (Landsat plus terrain data) images, and color composites as well as quantitative information were generated. Selected results of this preliminary investigation confirm the usefulness of the principal component analysis in a qualitative presentation of the multi-band data and its association with a significant reduction in dimensionality. However, unlike some other investigators, we found that the full dimensionality must be retained when the information content of the data has to be preserved quantitatively.

A program system for efficient multispectral classification

Pixelwise multispectral classification is an important tool for analyzing remotely sensed imagery data. The computing time for performing this analysis becomes significantly large when large, multilayer images are analyzed. In the classical implementation of the supervised multispectral classification assuming gaussian-shaped multidimensional class-clusters, the computing time is furthermore approximately proportional to the square of the number of image layers. This leads to very appreciable CPU-times when large numbers of multispectral channels are used and/or temporal classification is performed. In order to decrease computer time, a classification program system has been implemented which has the following characteristics: (1) a simple one-dimensional box classifier, (2) a multidimensional box classifier, (3) a class-pivotal “canonical” classifier utilizing full maximum likelihood and making full use of within-class and between-class statistical characteristics, (4) a hybrid classifier (2 and 3 combined), and (5) a local neighbourhood filtering algorithm producing generalized classification results. The heart of the classifier is the class-pivotal canonical classifier. This algorithm is based upon an idea of Dye suggesting the use of linear transformations making possible a simultaneous evaluation of a measure of the pixel being likely not to belong to the candidate class as well as computing its full maximum likelihood ratio. In case it is more likely to be misclassified the full maximum likelihood evaluation can be truncated almost immediately, i.e. the candidate class can often be rejected using only one or two of the available transformed spectral features. The result of this is a classifier with CPU-time which is empirically shown to be linearly dependent upon the number of image layers. The use of the hybrid classifier lowers the CPU-time with another factor of 3–4. Furthermore, for certain problems like classifying water-non water a single spectral band is often sufficient for making decisions and for such cases the simple one-dimensional box is used. The program constitutes a very efficient and flexible tool for multispectral classification. In many cases, the result from the pixelwise classifier is only the first step in the analysis procedure and often one wants a generalized classified image. The CLEANUP-routine (5) is a tool for achieving this by comparing the class-assignment for a particular pixel with those of the surrounding pixels. This scheme for supervised classification has been implemented and incorporated in several program systems for multilayer image processing and analysis like the batch-mode system on IBM 370 (TGLIB), the dialog system for batch-job summission on IBM 370 (PICCOLA) both in Stockholm as well as the dialog system on DEC 10 PILIP (PIXLIB) and its interactive supplement on PDP 11 34 (PIXLIB) at FOA 3 in Linköping. This work was partly financed by, and PICCOLA is now administered under, DFR (Swedish Board for Space Activity).

The logic of multispectral classification and mapping of land

The use of multispectral reflectance data as surrogates for land attributes must be done within strict rules of logic and with a recognition of judgmental factors such as the use of a priori or a posteriori classification schemes. The naming and describing of spectral classes as surrogates of information classes is a critical element in the logic of mapping and must be complete and logically consistent. Maps of information classes derived from multispectral data should be portrayed without class boundaries so as to indicate the degree of homogeneity or heterogeneity of classes on maps in which these characteristics are of major importance.

Evaluation of LANDSAT image data for land-use mapping

A computer aided extraction of land-use information from LANDSAT image data, gathered on 9 August 1975, was conducted. Specifically the study was designed to develop a method for supervised multispectral classification of LANDSAT data of a central European region and to verify the classification results on a pixel-by-pixel basis and by comparison of the land-use inventory with aerial photo interpretation in test areas. The method enables the integration and processing of additional data. For the test area Speyer in the Rhine valley with small size allotments and a variety of land-use it was found that the pixel-by-pixel comparison is rather difficult and has a high potential failure rate. Seven classes were considered. With the integration of subclasses to main classes higher classification accuracy can be reached. The highest classification accuracy could be obtained for agricultural land (97%); orchards, pasture (91%) and coniferous forest (84%) in the land-use inventory of Speyer. The study confirms that the degree of fit rises with the size of homogeneous land-use zones. The study results yield that for larger areas of homogeneous land-use in central Europe this method will reach acceptable classification accuracies for land-use inventories. The land-use information can be presented in colour coded thematic maps with the same scale and projection of existing maps.

A Relaxation Method for Multispectral Pixel Classification

Three approaches to reducing errors in multispectral pixel classification were compared: 1) postprocessing (iterated reclassification based on comparison with the neighbors' classes); 2) preprocessing (iterated smoothing, by averaging with selected neighbors, prior to classification); and 3) relaxation (probabilistic classification followed by iterative probability adjustment). In experiments using a color image of a house, the relaxation approach gave markedly superior performance; relaxation eliminated 4-8 times as many errors as the other methods did.

A mixed integer programming approach to multi-spectral image classification

A supervised discriminant mixed integer programming algorithm (DISMIP) is described which achieves either linear or non-linear separation, without assuming any specific probability distribution. This system offers greater flexibility in dealing with problems of multi-spectral classification. If the training sets are disjoint, a strictly separating surface is generated that maximizes a “dead zone” between the sets. If the sets intersect, a surface is generated that minimizes a specified misclassification error. The system has been experimentally tested in three practical applications and the results are given in comparison with a supervised classification using the LARSIS classifier. (1)