ASPRS

PE&RS May 1997

VOLUME 63, NUMBER 5
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
JOURNAL OF THE AMERICAN SOCIETY FOR PHOTOGRAMMETRY AND REMOTE SENSING

Peer Reviewed Articles

485-491 Rule-Based Classification of Water in Landsat MSS Images Using the Variance Filter
Paul A. Wilson

Abstract
Download Full Article
The variance filter is a textural algorithm capable of distinguishing between flat, uniform water bodies and cloud or mountain shadows when applied to satellite imagery. The filter output forms the basis of rules used by a knowledge-based classifier, which segments water adaptively. The use of the filter in the unsupervised classification of water is demonstrated on two spectrally varied Landsat MSS images. The same images are segmented using a conventional thresholding algorithm. The two algorithms identify a similar proportion of the water pixels in both images; however, the rules-based algorithm does not generate any false positives, whereas the threshold algorithm misclassifies many shadow pixels as water. The rules-based algorithm is less efficient at finding small lagoons and swamps than at finding large water bodies.

Detecting Subpixel Woody Vegetation in Digital Imagery Using Two Artificial Intelligence Approaches
Patricia G. Foschi and Deborah K. Smith

Abstract
Download Full Article
Small strips or patches of woody vegetation, typical landscape elements in many farming areas, are frequently not detected by standard computer-assisted clessification of digital satellite imagery because such landscape elements are smaller than the pixel size and are mixed with other classes. This study essentially compares two artificial intelligence opproaches - machine-vision and neural-network rnethods - developed to improve classification accuracy for this mixed pixel problem. Simulated multispectral and panchromatic SPOT HRV imagery of lowland Britain was used to test both methods. Compared to standard supervised multispectral classification, both methods yield significant improvements in detecting subpixel woody vegetation. In general. the machine-vision approach outperformed the neural-netwark approach. However, because each method generated different types of misclassifications, a classification map representing only the woody vegetation found by both methods provided the results with the least amount of overall error. 

Performance of a Neural Network: Mapping Forests Using GIS and Remotely Sensed Data
A.K. Skidmore, B.J. Turner, W. Brinkhof, and E. Knowles

Abstract
Download Full Article
Neural networks have been proposed to classify remotely sensed and ancilliary GIS data, In this poper, the back propagation algonithm is critically evaluated using as an example, the mapping of a eucalypt forest on the far south coast of New South Wales, Australia. A GIS database was combined with Landsat thematic mapper data, and 190 plots were field sampled in order to train the neural network model and to evaluate the resulting classifications. The results show that the neural network did not accurately classify GIS and remotely sensed data at the forest type level [Anderson level III], though conventional classifiers also perform poorly with thiis type of problem. Previous studies using neural networks have classified more general (e.g., Anderson Level I, II) landcover types at a higher accuracy than those obtained here, but mapped land cover into more general themes. Given the poor classification results ond the difficulties associated with the setting up of suitable parameters for the neural-network (backpropagation) algorithm, it is concluded that the neural network approach does not offer significant advantages over conventional classification schemes for mapping eucelyptus forests from Landsat TM and ancillary GIS data at the Anderson Level III forest type level.

Texture Analysis of Tropical Rain Forest Infrared Satellite Images
Robert Riou and Frederique Seyler

Abstract
Download Full Article
This paper describes a new application of the Fourier transform for texture analysis of low reflectance, rain forest satellite images. After correction of edge effects, angular power spectra have been analyzed on windows of 64 by 64 pixels in size. Results of spectra analysis are compared to visually extracted linear features of near infrared SPOT images from South Cameroon forested areas. This method allows for the detection of low-contrast, periodic luminance variations and the extraction of lineament direction with an accuracy of 100. An algorithm of processing of the whole image aimed at regional studies of texture networks is proposed. Applications to structural geology, forest-type discrimination, and soil studies in the tropics are briefly evaluated.

523-533 Multisource Classification of Complex Rural Areas by Statistical and Neural-Network Approaches
L. Bruzzone, C. Conese, F. Maselli, and F. Roli

Abstract
Download Full Article
The usefulness of spectral (Landsat-5 Thematic Mapper images), texture (grey-level co-occurrence matrix statistics), and ancillary (terrain elevation, slope, and aspect) data to characterize two complex rural areas in central Italy is quantitatively demonstrated. A statistical and a neural- network classification approach are applied to such a multisource data set, and their classification performances are assessed and compared. The classification performances of the two approaches are quantitatively evaluated in terms of global and conditional Kappa accuracies. The Zeta statistics is used to evaluate the statistical significance of the different classification accuracies obtained by the two approaches by using multisource data.

535-544 The Effect of Neural-Network Structure on a Multispectral Land-Use/Land-Cover Classification
Justin D. Paola and Robert A. Schowengerdt

Abstract
Download Full Article
While neural networks are now an accepted alternative to statistical multispectral classification techniques for remote sensing image classification, the network approach presents both unique challenges and abilities. The size of the hidden layer must be determined by trial and error, and the random initial weight settings result in different paths for the training procedure, making the network a non-deterministic classifier. For the sample classification presented here, it was found that there was a range of optimal hidden layer sizes below which the occupancy decreased and above which the training time increased. However, it was also found that, for a fairly wide range, the hidden layer size made little difference to the final classification accuracy. Initial weight randomization was as much of a factor as hidden layer size. Using 3 by 3 windows of data in each band was found, despite increased training time per iteration, to achieve similar accuracy with less overall training time, although with less consistency.
Top Home