Remote Sensing & GIS 2 Flashcards
For image histogram (a) below, chose one contrast enhancement technique that you think would best enhance the image.
Describe the technique and explain the reason for your choice.
Sketch the histogram and draw the shape of the function in the relevant histogram.
Linear contrast enhancement (LCE).
Here, because the image histogram is normally distibuted but the output DN range is wider than the input DN range, the function improves image contrast without distorting the image information.
I.e. the LCE does nothing but widen the increment of DN levels and shift histogram position along the image DN axis.
Linearly scale the DN range of the image to the full dynamic range of a display system (8 bits) based on the maximum and minimum of the input image X.
For image histogram (b) below, chose one contrast enhancement technique that you think would best enhance the image.
Describe the technique and explain the reason for your choice.
Sketch the histogram and draw the shape of the function in the relevant histogram.
Logarithmic contrast enhancement.
Histogram (b) is in a form of a logarithmic normal distribution. Hence, a logarithmic function will modify such a histogram into a shape of normal distribution.
The gradient of the function is greater than 1 in the low DN range thus it spreads out low DN values while in the high DN range, the gradient of the function is less than 1 and thus it compresses high DN values.
As the result, logarithmic contrast enhancement shifts the peak of image histogram to the right and highlights the details in dark areas in an input image.
For image histogram (c) below, chose one contrast enhancement technique that you think would best enhance the image.
Describe the technique and explain the reason for your choice.
Sketch the histogram and draw the shape of the function in the relevant histogram.
Exponential contrast enhancement.
Histogram (c) is in a form of an exponential normal distribution. Hence, an exponential function will modify such a histogram into a shape of normal distribution.
The gradient of the function is less than 1 in the low DN range thus it compresses low DN values while in the high DN range, the gradient of the function is greater than 1 and thus it spreads out high DN values.
As the result, logarithmic contrast enhancement shifts the peak of image histogram to the left and enhances details in light areas at the cost of suppressing the tone variation in dark areas.
Exponential and logarithmic functions are the inverse of each other.
What is the preferred technique for generating a colour composite image with optimised colour balance and contrast?
How is this done?
Balance contrast enhancement technique (BCET)
Colour bias is one of the major causes of poor colour composite images. The BCET is a simple solution to this problem, where the three bands used for colour composition must have an equal value range and mean.
Using the parabolic function [y = a(x - b)2 + c] derived from an input image, BCET can stretch (or compress) the image to a given value range and mean without changing the basic shape of the image histogram.
Thus three image bands for colour composition can be adjusted to the same value range and mean to achieve a balanced colour composite.
Use a diagram for to explain the principle of principal component analysis (PCA).
- In general, for multi-spectral imagery, the reflective bands are highly correlated, and, for example, if there is a correlation of 0.99 between two bands, this means that there is 99% information redundancy between the two bands, with only 1% of unique information.
- As such, multi-spectral imagery is not efficient for information storage.
- Consider an m band multi-spectral image as an m-dimensional raster dataset in an m dimension orthogonal coordinate system, forming an m dimensional ellipsoid cluster (represented by its covarience matrix, Σ<em>x</em>).
- Then the coordinate system is oblique to the axes of the ellipsoid data cluster if the image bands are correlated (e.g. correlation > 0.8 is very high).
- The axes of the data ellipsoid cluster formulate an orthogonal coordinate system and in this system, the same imagery data are represented by n (n<=m) independent components that are called principal components.
- In other words, the principal components are the imagery data representation in the axes of the ellipsoid data cluster.
- Thus, principal component analysis is a coordinate rotational operation to rotate the coordinate system of the original image bands to match the axes of the ellipsoid of the imagery data cluster.
- The first PC is represented by the longest axis of the data cluster and the second PC the second longest, and so on.
- The axes representing high order PCs may be too short to represent any substantial information and then the apparent m dimensional ellipsoid is effectively degraded to n (n<=m) independent dimensions.
- In this way, PCA reduces image dimensionality and represent the nearly same imagery information with fewer independent dimensions in a smaller dataset without redundancy.
Summarize the principle of principal component analysis (PCA).
- The principal component analysis is a linear transformation converting m correlated dimensions to n (n<=m) independent (uncorrelated) dimensions.
- This transform is a coordinate rotation operation to rotate the coordinate system of the original image bands to match the axes of the ellipsoid of the imagery data cluster.
Describe the principle and operation of PCA based decorrelation stretch.
The PCA based decorrelation stretch generates a colour composite from three image bands with reduced inter-band correlation and thus more distinctive and saturated colours without distortion in hues.
The idea of PCADS is to stretch multidimensional imagery data along their PC axes (the axes of the data ellipsoid cluster) rather than original axes representing image bands.
In this way, the volume of data cluster can be effectively increased and the inter-band correlation is reduced as illustrated below:
Describe the major steps of PCADS.
The PCADS is achieved in three steps:
- PCA to transform data from the original image bands to PCs.
- Contrast enhancement on each of the PCs (stretching the data cluster along PC axes).
- Inverse PCA to convert the enhanced PCs back to the relevant image bands.
Compare the images of band average, intensity of HSI, and PC1, and discuss their similarity (in operation) and common merits, and comment on the differences between them.
Average image, intensity image of HSI and PC1 image are largely in common.
The three types of images are all summation of spectral bands and they all increase the image SNR.
- Band average is an equal weight summation of any number of image bands,
- Intensity image of HSI is an average of three bands used for RGB HSI transformation,
- PC1 is a weighted summation of all the image bands based on the 1st eigenvector of the image covariance matrix.
Compare the three DS techniques, PCADS, HSIDS, DDS, in principle, results and processing efficiency.
- The PCADS technique is similar to the HSIDS in results though based on quite different principles.
- PCADS is statistically scene dependent as the whole operation starts from the image covariance matrix. It can be operated on all image bands by one go.
- The HSIDS is not statistically scene dependent and only operates on three bands.
- Both techniques involve complicated forward and inverse coordinate transformations.
- In particular, the PCADS requires quite complicated inverse operations of eigenvector matrix for inverse PC transformation and therefore not widely used.
- The Direct Decorrelation Stretch (DDS) is the most efficient technique and it can be quantitatively controlled based on the saturation level of the image.
Using diagrams, explain the principle of additive RGB colour display based on Tristimulus Colour theory. Discuss how colours are used as a means to visualise information beyond the spectral range of human vision
- The human retina has 3 types of cones. The response of each type of cone is a function of the wavelength of the incident light and peaks at 440nm (B), 545nm (G), and 680nm (R) individually.
- I.e. each type of cone is primarily sensitive to one of the primary colours: B, G or R.
- A colour perceived by a person depends on the proportion of each of these three types of cones been stimulated and thus can be expressed as a triplet of numbers (r, g, b) even though visible light is electromagnetic radiation in a continuous spectrum of 400-700nm.
- Digital image colour display entirely based on the tristimulus colour theory.
- A colour monitor, like a colour TV, is composed of three geometrically registered guns: R, G and B.
- In the red gun, pixels of an image are displayed in reds of different intensity (i.e. dark red, light red etc.) depending on their DNs. So are the green and blue guns.
- Thus, if three bands of a multi-spectral image are displayed in R, G and B simultaneously, a colour image display is generated, in which the colour of a pixel is decided by its DNs in R, G and B bands (r, g, b).
- This kind of colour display system is called Additive RGB Colour Composite System.
- In this system, different colours are generated by additive combinations of R, G and B components.
Describe the pseudo colour display method.
- Human eye can recognize far more colours than grey levels.
- Thus colour may be used very effectively to enhance small grey level differences in a monochrome image.
- The technique to display a monochrome image as a colour image is called pseudo colour display.
- A pseudo colour image is generated by assigning each grey level to a unique colour.
- This can be done by interactive colour editing or by automatic transformation based on certain logic.
Discuss the advantages and disadvantages of the pseudo colour display method in image visualisation and interpretation.
- The advantage of psuedo colour display is also its disadvantage.
- When a digital image is displayed in a grey scale based on its DNs in a B/W display, the quantitative sequential relationship between different DN values is effectively presented.
- This crucial information is lost in a pseudo colour display bc the assigned colours to various grey levels are not quantitative and sequential.
- Indeed, the image of a pseudo display is an image of symbols rather than numbers; it is no longer a digital image.
- We can regard the grey B/W display as a special case of pseudo colour display in which a sequential grey scale based on DN levels is used instead of a colour scheme.
What is meant by smoothing in digital image filtering?
- Smoothing (low pass) filters are designed to remove high frequency information and retain low frequency information,
- thus, reducing the noise at the cost of degrading details in an image
- The figure illustrates a typical low pass (smoothing) filter H(u,v) and the corresponding PSF h(x,y).
- Most kernel filters for smoothing involve weighted average among the pixels within the kernel.
- The larger is the kernel, the lower frequency information is retained.
Give some examples of low pass filter kernels.
- Mean filters
- Weighted mean filters
- Gaussian filter
- Edge preserve low pass filters
Discuss the major drawback of mean filters and the importance of edge preserve smoothing filters.
- Smoothing based on average is effective to eliminate noise pixels which are often distinct as very different DNs from their surrounding pixels, but the process blurs an image by removing high frequency information.
- For this reason, “edge-preserving smoothing” technique becomes an important research topic of filtering.
Describe the concept of k nearest mean filter and discuss its applications and merits.
- Kind of Edge preserve low pass filter
- Re-assign a pixel xij (central pixel) of an image X to the average of the k neighbor pixels in the kernel window whose DNs are closest to that of xij.
- A typical value of k is 5 for a 3x3 square window.
- In the image on the LHS, 0 is noise bc it’s significantly different from the neighborhood pixels.
- There is a distinct image boundary in the image on the RHS that is preserved after filtering.
Describe the concept of median filter and discuss its applications and merits.
- Kind of Edge preserve low pass filter
- Re-assign a pixel xij of image X to the median DN of its neighboring pixels (including itself) in a kernal window (e.g. 3x3)
Describe the concept of adaptive median filter and discuss its applications and merits.
- Kind of Edge preserve low pass filter
- The adaptive median filter is designed based on the basic principle of median filter with such an approach:
Describe the concept of majority filter and discuss its applications and merits.
- Kind of Edge preserve low pass filter
- This is a rather democratic filter.
- A pixel is re-assigned to the most popular DN among its neighbourhood pixels.
- This filter performs smoothing based on the counting of pixels in the kernel rather than numerical calculations.
- Thus it is suitable for smoothing images of non-sequential data (symbols) such as classification images.
- For a 3x3 kernel, the recommended majority number is 5.
- If there is no majority found within a kernel window, then the central pixel in the window remains unchanged.
To smooth a classification image, what filter is appropriate and why? Describe this filter with an example.
- There are 6 pixels with DN=6, t.f. the central DN, 2, is replaced by 6.
- For a classification image, the numbers in this window are the class numbers and their meaning is no difference to class symbols A, B and C.
- If we use a mean filter, the average of the DNs in the window is 5.3. Class 5.3 has no meaning in a classification image.
What is meant by edge enhancement in digital image filtering?
- Edges and textures in an image are typically high frequency information.
- High pass (edge enhancement) filters remove low frequency image information and t.f. enhance high frequency information such as edges.
- The figure below illustrates a typical high pass filter H(u,v) and the corresponding PSF h(x,y).
Compare Laplacian filters with gradient filters and discuss their applications
- Most commonly used edge enhancement filters are based on first and second derivatives or Gradient and Laplacian.
- The two types of high pass filters work in different ways:
- Gradient is the first derivative at pixel f(x,y) and as a measurement of DN change rate, it is a vector characterising the maximum magnitude and direction of the DN slope around the pixel.
- Laplacian, as the second derivative of f(x,y), is a scalar that measures the change rate of gradient.
- I.e. Laplacian describes the curvature of a slope but not its magnitude and direction.
What does this figure show?
Explain the figure in the context of edge enhancement.
- Geometric meaning of first and second derivatives.
- A flat DN slope has a constant gradient but zero Laplacian bc the change rate of a flat slope is zero.
- For a slope with constant curvature (an arc of a circle), the gradient is a variable while the Laplacian is a constant.
- Only for a slope with varying curvature, both gradient and Laplacian are variables.
- This is why Laplacian suppresses all the image features except sharp edges where DN gradient changes dramatically, while gradient retains edge as well as slope information.
Give the mathematical definitions of gradient and Laplacian filters.
Give examples of gradient filters.
- Prewitt filters
- Sobel filters
Give examples of Laplacian filters.
- Standard Laplacian filter
- Laplacian filter
- Edge sharpening filter
What is an edge sharpening filter?
What are the major applications of edge sharpening filter?
What is the purpose of edge preserved filters?
In order to remove random noise with the minimum degradation of resolution, various edge preserved filters have been developed such as adaptive median filter.
Discuss the principle of SPCA for colour composition.
(Not actually in q’s, might be useful for understanding)
- As PCs are independent without information redundancy, colour composites of PCs are often very effective to highlight particular ground objects and minerals that are not distinguishable in colour composites of the original bands.
- However, in PC colour composites, the noise may be exaggerated because the high rank PCs contain significantly less information than lower rank PCs and have very low SNR.
- When PC images are stretched and displayed in the same value range, the noise in higher rank PCs is improperly enhanced.
- We would like to use three PCs with comparable information levels for colour composite generation.
- Selective Principal Component Analysis (SPCA) techniques can produce PC colour composites condensing maximum information of either topographic or spectral features and meantime makes the information content in each PC displayed in RGB better balanced.
- There are two types of SPCA: dimensionality and colour confusion reduction and spectral contrast mapping.
Discuss the SPCA technique dimensionality and colour confusion reduction and its application.
- The spectral bands of a multi‐spectral image are arranged into three groups and each group is composed of highly correlated bands.
- PCA is performed on each group and then the three PC1s derived from these three groups are used to generate an RGB colour composite.
- As bands in each group are highly correlated, the PC1 concentrates the maximum information of each group.
- For six reflective spectral bands of a TM or ETM+ image, the technique may condense more than 98% variance (information) in the derived PC1 colour composite. The recommended groups for six reflective spectral bands of TM or ETM+ images are shown in the Table.
- The approach is actually equivalent to generating a colour composite using broader spectral bands, given that a PC1 is a positively weighted summation of the bands involved in the PCA.
- It essentially creates a colour composite of a broad visible band in blue, an NIR band in green and broad SWIR band in red.
Discuss the SPCA technique dimensionality and colour confusion reduction and its application.
- The spectral bands of a multi‐spectral image are arranged into three groups and each group is composed of highly correlated bands.
- PCA is performed on each group and then the three PC1s derived from these three groups are used to generate an RGB colour composite.
- As bands in each group are highly correlated, the PC1 concentrates the maximum information of each group.
- For six reflective spectral bands of a TM or ETM+ image, the technique may condense more than 98% variance (information) in the derived PC1 colour composite. The recommended groups for six reflective spectral bands of TM or ETM+ images are shown in the Table.
- The approach is actually equivalent to generating a colour composite using broader spectral bands, given that a PC1 is a positively weighted summation of the bands involved in the PCA.
- It essentially creates a colour composite of a broad visible band in blue, an NIR band in green and broad SWIR band in red.
Discuss the SPCA technique spectral contrast mapping and its application.
- A more interesting and useful approach (as opposed to DCCR) is spectral contrast mapping, where the primary objective is to map the contrast between different parts of the spectrum, and so to identify information unique to each band rather than information in common.
- For this purpose, PC2s derived from band pairs are used instead of PC1s.
- By using only two bands as inputs, the information that is common to both bands is mapped to PC1 and the unique information is mapped to PC2.
- In general, low or medium correlation between the bands in each pair is preferred for this approach.
- The recommended grouping for the six reflective spectral bands of TM/ETM+ is listed in the Table.
- Based on the above principle, groups of three or more bands may also be used for spectral contrast mapping.
- The technique can generate spectrally informative colour composites with significantly reduced topographic shadow effects and better SNR.
Describe the Decorrelation Stretch (DS) technique, PCADS, in principle, results and processing efficiency.
- The idea of PCADS is to stretch multidimensional imagery data along their PC axes (the axes of the data ellipsoid cluster) rather than original axes representing image bands.
- In this way, the volume of data cluster can be effectively increased and the inter-band correlation is reduced.
- The PCADS technique is similar to the HSIDS in results (both increase the three dimensionality of the data cluster) though based on quite different principles.
- PCADS is statistically scene dependent as the whole operation starts from the image covarience matrix.
- It can be operated on all image bands by one go.
- Both PCADS and HSIDS involve complicated forward and inverse coordinate transformations.
- In particular, the PCADS requires quite complicated inverse operations of eigenvector matrix for inverse PC transformation and t.f. not widely used.
Comment on the Decorrelation Stretch (DS) technique, HSIDS, in principle, results and processing efficiency.
- The HSIDS technique is similar to the PCADS in results (both increase the three dimensionality of the data cluster) though based on quite different principles.
- The HSIDS is not statistically scene dependent and only operates on three bands.
- Both HSIDS and PCADS involve complicated forward and inverse coordinate transformations.
Comment on the Decorrelation Stretch (DS) technique, DDS, in principle, results and processing efficiency.
The Direct Decorrelation Stretch (DDS) is the most efficient technique and it can be quantitatively controlled based on the saturation level of the image.
What are the major advantages and disadvantages of PC colour composition?
- As PC1 is mainly topography, colour composites excluding PC1 may better present spectral information with topography subdued.
- PCs represent condensed and independent image information and therefore produce more colourful (i.e. informative) colour composites.
- However, here we have a problem in that a PC is a linear combination of the original spectral bands, so its relationship to the original spectral signatures of targets representing ground objects is no longer apparent.
- To solve this problem, a feature‐oriented PC selection (FPCS) method for colour composition was proposed.
Use a diagram for a two dimensional case to explain the principle of principal component analysis (PCA).
- In general, for multi-spectral imagery, the reflective bands are highly correlated, and, for example, if there is a correlation of 0.99 between two bands, this means that there is 99% information redundancy between the two bands, with only 1% of unique information.
- As such, multi-spectral imagery is not efficient for information storage.
- As shown by the 2D illustration, suppose the image data points form an elliptic cluster;
- The aim of PCA is to rotate the orthogonal co‐ordinate system of band 1 and band 2 to match the two axes of the ellipsoid, the PC1 and PC2.
- The coordinates of each data point in the PC co-ordinate system will be the DNs of the corresponding pixels in the PC images.
- The first PC is represented by the longest axis of the data cluster and the second PC the second longest, and so on.
- The axes representing high order PCs may be too short to represent any substantial information, and then the apparent m-dimensional ellipsoid is effectively degraded to n (n<=m) independent dimensions.
- In this way, PCA reduces image dimensionality and represent the nearly same imagery information with fewer independent dimensions in a smaller dataset without redundancy.
What is a point operation for image processing?
Give the mathematical definition.
A process that transforms a single input image, X, to a single output image, Y, through a function f, in such a way that the DN of an output pixel, yij, depends only on the DN of the corresponding input pixel xij.
yij = f(xij)
When X represents a digital image and xij, the DN of any pixel in the image at the ith line and jth column, Y represents the image derived from X by a function, f, and yij is the output value corresponding to xij.
Using a diagram, explain why point operation is also called histogram modification.
- Point operation is also called histogram modification because the operation only alters the histogram. h(x), of an image but not spatial relationship of image pixels.
- For pixels with the same input DN, x, but different locations, the function, f, will produce the same output DN, y.
- T.f. point operation can be more concisely defined as:
y = f(x)
What is contrast enhancement?
- Sometimes called radiometric enhancement or histogram modification, contrast enhancement is the most basic but very effective and efficient technique to optimise the image contrast and brightness for visualization or highlight information in particualr DN ranges.
- Contrast enhancement is a point operation that modifies the image brightness and contrast but does not alter the image size and textures.
What is a false colour composite?
If image bands displayed in red, green and blue do not match the spectra of these three primary colours, a false colour compositeimage is produced.
A typical example is the so called standard false colour compositein which nearer infrared band is displayed in red, red band in green and green band in blue.