Image analysis – Applications – Target tracking or detecting
Reexamination Certificate
1999-02-19
2002-03-12
Boudreau, Leo (Department: 2621)
Image analysis
Applications
Target tracking or detecting
C348S144000
Reexamination Certificate
active
06356646
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Technical Field
This invention relates to the computer processing and display of digital, multispectral imagery for the purpose of identifying ground targets and classifying the imagery to create thematic maps.
2. Description of Prior Art
Ground targets in imagery have been identified commonly through the subjective process of an art referred to as photointerpretation. This process relies on shapes, tones, textures, colors, and associations in the image to infer the nature of the ground feature. The results are highly correlated with the skill and experience of the photointerpreter.
Multispectral imagery is a collection of coregistered images (typically <2
3
), each image being collected in a different, broad region of the electromagnetic spectrum commonly called a band. These bands are collected nominally between 0.400 &mgr;m (micrometers) and 2.500 &mgr;m wavelength; this region is generally referred to as the visible/infrared-red (VIR) region. The computer classification of multi-spectral imagery, to create thematic maps, has usually been of two general types, referred to as supervised and unsupervised.
In the case of supervised classification, regions within the image, that are believed to represent relatively homogeneous clusters of pixels (picture elements) that are characteristic of a particular type of ground cover of interest, are selected to be what are known as training sites. The selection can and most often does proceed by an experienced photointerpreter marking the boundaries of the training site with an on-screen cursor controlled with some pointing device such as a mouse, trackball, joystick, lightpen, or keyboard arrow-keys. The selection of particular training-site locations within an image, and their resultant shapes, is subjectively determined by an operator. The selection is based on their interpretation of the image and possibly is guided by information called “ground truth” obtained from on-site reconnaissance on the ground.
These training sites are then used by various computer algorithms, again commonly selected by an operator from a plurality of choices, and typically statistical in nature, (such as the Maximum Likelihood classifier) to sequentially test the spectral characteristics of every pixel in the image against the pixels in the selected training sites. The tests performed are for a similarity metric for each pixel with a plurality of multispectral bands that may represent all bands recorded or a sub-set selected by the operator. The definition of the agreement or similarity varies with the algorithm selected. The intent is to categorize or classify every pixel as being most like the pixels in a particular training site, when considered from the perspective of the n-dimensional feature space defined by the number of bands selected for the comparison.
In the unsupervised classification approach, an operator selects a particular algorithm from a plurality of choices commonly provided by the manufacturer of the image processing software. That algorithm then performs iterative comparisons in n-dimensional space, where n is the number of multispectral bands chosen by the operator to be considered. The comparisons result in every pixel being assigned to a cluster of spectrally similar pixels. If the statistical characteristics of any or all of the existing clusters exceed certain thresholds, either hardcoded by the software manufacturer or selected by the operator, clusters can be split or lumped together. After a pre-selected number of iterations, or alternatively a pre-selected number of clusters is attained, the program halts and displays the spectrally dissimilar clusters, wherein the clusters are commonly pseudo-colored to help in their visual differentiation. The operator must then subjectively decide what every cluster represents in terms of a ground target or type of ground cover. This is again accomplished through either photointerpretation procedures, or the collection of ground-truth from onsite inspection, or a combination of both.
Numerous band-ratios and indices, utilizing two bands, have been in use for years to accentuate spectral features. Thresholding a band-ratio (or vegetation index) can provide a binarized (two-level) classification of the images. Where to establish the threshold is subjective.
It can be seen that the subjective judgement of the operator is extremely important in obtaining accurate results for all the methods discussed above. Their experience is extremely important in making the subjective decisions. In addition, the availability of accurate ground truth is important in achieving good classification results. Although the algorithmic processing of the imagery can be performed for an n-dimensional feature space where n is greater than 1, the operator is constrained to only being able to see the effects of three bands at any one time. This affects the ability to either correctly select homogeneous training sites or properly identify the corresponding type of ground cover for a particular cluster-class derived from unsupervised classification. One does not know a priori what the best three multispectral bands are to help identify the various types of ground cover in the image.
An additional problem is that one can expect that every algorithm chosen will result in a classification of the ground cover that is different from the results of every other algorithm. The initial selection of an appropriate algorithm from the plurality of choices available will be influenced by the experience of the operator. The acceptance or rejection of the classification results will also be a subjective decision made by the operator. While there are objective tests of the accuracy of classification, even the interpretation of the tests is ultimately subjective.
A common problem in multispectral classification, particularly with high spatial-resolution imagery, is the incorporation and ambiguous identification of a shadow class. Related to this is, if the operator does not recognize the existence of one or more spectrally unique types of ground cover, those classes will be relegated to an ‘unknown’ or ‘undefined’ class with supervised classification.
Another approach for creating thematic maps, which has evolved in recent years, is an extension of multispectral remote sensing. If a sufficient number of as discrete bands of the electromagnetic spectrum are sampled, the resulting plot of the apparent reflectance of a ground target approaches a continuous, smooth curve. The imaging instruments that collect a very large number of bands, usually between about 2
6
to 2
8
narrow bands, nominally between 0.400 &mgr;m and 2.500 &mgr;m wavelength, are called hyperspectral sensors. The advantage of hyperspectral imagery is that, with careful calibration, correction of illumination variations, and compensation for atmospheric absorption and scattering effects, an apparent reflectance can be derived from the radiance values for every pixel; then the pixels can be compared to a library of laboratory-derived reflectance spectra for a match. Thereby, ideally, the need for subjective judgement on the part of the operator is eliminated. However, the atmospheric corrections are not a trivial undertaking, the available spectral libraries are reasonably complete for minerals only, and the hyperspectral sensors are not extensively deployed as yet. The analysis and subsequent processing of hyperspectral imagery to create thematic maps is generally more difficult and requires a higher level of skill and experience than for multispectral imagery.
Multispectral (and hyperspectral) images are invariably displayed as a single false-color image created from a combination of three particular bands out of the total number of bands recorded. However, most remote sensing image processing software currently available only provides the ability to plot scattergrams from two bands simultaneously.
A ternary diagram (a.k.a. triangular plot) is a graphical plot based on the use of an equilateral triangle; it is used for displaying the relatio
Boudreau Leo
Tabatabai Abolfazl
LandOfFree
Method for creating thematic maps using segmentation of... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for creating thematic maps using segmentation of..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for creating thematic maps using segmentation of... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2820065