Basics of image processing

What will you learn?

importing image data to your RS processing/GIS software radiometric image correction geometric image correction (georeferencing and geo-coding) basic image enhancement contrast enhancement spatial enhancement single band image classification density slicing

RS data characteristics
RS data are not photos - they are images at best each image element (pixel) is described by a meaningful digital number (DN) DN reflects radiation value measured by a radiometer data files are binary, i.e. we cannot see DNs using a normal ASCII file editor DN for a single pixel can be written using 8 to 24 bits depending on range of DNs obtained from the radiometer we do image processing on DNs, but translate them into colors for visual interpretation RS data comes in various formats, not always compatible with the format your RS processing software uses 3

RS data formats

Line 1 Line 2 Line 3 .. .. .. .. .. .. .. .. .. .. .. Line n

single band file

BIL BSQ (band interleaved by line) (band sequential) Usually, there are sequential pixel data only. No information on row and column, no information when row changes. Header is sometimes an independent file.

Importing RS data
importing usually done by build in filters
if we know the raw file format, and if there is a proper filter in our RS software

importing single band general raster

you have to know: number of rows, number of columns, number of bytes per pixel best check: nrow * ncol * nbytes per pixel = size of the file in bytes

importing multiple band general raster

have to know: number of rows, number of columns, number of bytes per pixel, number of bands, interleaving format best check: nrow * ncol * nbytes per pixel * nbands = size of the file in bytes interleaving formats: BIL (band interleaved by line), BSQ (band sequential), BIP (pixel interleaved)

Displaying RS data
RS data are Digital Numbers, one band is usually written using 1 byte per pixel, i.e. values between 0 and 255 monitors usually display 256 shades of grey (i.e. different levels of brightness of white) In the process of displaying a single band image, brightness of the screen pixel corresponding to the image pixel is adjusted according to its DN value

255
232
209
185
162
139
116
93
70
46
23
0
6

Displaying RS data
if DNs have wider range than 0-255, or if different colour scheme than greyscale is to be used - LOOKUP TABLE (or representation) is applied Lookup table tells the computer, what colour to use to represent given DN value

255
232
209
185
162
139
116
93
70
46
23
0

Multiple band images are displayed in a different way. More about it later.

Radiometric corrections

there is a difference between top of the atmosphere (TOA) radiation and ground level radiation - due to absorption and scattering in the atmosphere

sensors are sometimes faulty/, what results in dropped lines or pixels TM data often exhibits sixth-line striping resulting from difference in sensitivities of sensors in the array

radiometric correction procedure eliminates these effects

Radiometric corrections
Striping

DNs of bad lines are adjusted so as histogram of the bad line matches that of the good lines 9

Radiometric corrections
Dropped lines

DNs of bad lines are obtained as average of the neighbouring DNs 10

Radiometric corrections
Atmospheric effects correction

Atmospheric scattering causes an addititve effect, i.e. TOA radiation is higher than at ground level Shorter wavelenghts are affected stronger

TM blue band

TM green band Methods dark pixel subtraction radiance to reflectance conversion linear regressions between bands atmospheric modelling Important when quantitative calculations are performed, can be ignored for most qualitative applications 11

Georeferencing and geometric corrections

RS data has geographical meaning (i.e. pixels correspond to a given place on Earth) geometry of the raw image does correspond neither to the real geometry of the area nor to that used to represent that area in maps

Aim of georeferencing: to make RS data compatible with other spatial data (maps, ground measurements etc.) in a GIS system

12

Georeferencing and geometric corrections

RS data are affected by geometric distortions due to
sensor geometry, platform instabilities, earth rotation, earth curvature etc.

Some of these distortions are corrected by image supplier, others have to be corrected by referencing images to exisiting maps or other images.

13

Georeferencing and geometric correction

image georeferencing or calibration
defining geographic coordinates without changing the geometry and DNs of the image

image geo-coding or geometric correction

adjusting the geometry of the image to a given geometry, may involve change of DNs

14

Image georeferencing
(a.k.a calibration, registration, rectification)

defining (geographic) coordinates for each pixel in the image two steps:
attaching coordinates to a few known pixels in the image defining the formula (coordinate transform) that allows for calculating coordinates for the remaining pixels

xcoord,ycoord = f(row,col)

15

Image georeferencing
(a.k.a calibration, registration, rectification)
attaching coordinates to a few points in the image georeferencing corners (linear transformation only)
coordinates known from satellite position data and geometry of the imaging system, and data has been corrected already for some geometry errors, coordinates of corners often included in image header (e.g. for TM images)

16

Image georeferencing
(a.k.a calibration, registration, rectification)
attaching coordinates to a few points in the image georeferencing using ground control points (GCP) (a.k.a image-to-map registration) georeferencing using image-to-image registration

georeferenced image raw image

http://glcf.umiacs.umd.edu tip: GCPs should be distributed evenly throughout the image. For a TM scene - 15-30 points. 17

Image georeferencing
Defining the coordinate transform affine (linear): polynomial: xcoord = A * row + B * col + C xcoord = A * row2 + B * row + C * col2 + D * col + E * row * col + F

raw image pixels

raw image pixels

geographic coordinates

affine (linear)

polynomial 18

tip: for satellite images - degree not higher than 2, preferably 1

Image geometric correction

(a.k.a geo-coding)

identification of target grid (pixel size, image extents and projection) defining resampling method:
nearest neighbour - for images bilinear and bicubic - for DEM data, but not for images, because it changes DNs

nearest neighbour

bilinear interpolation

19

Image enhancement

Aim: making an image better interpretable for a particular application contrast enhancement
linear stretching histogram equalization piecewise stretching

spatial enhancement
low pass filter high pass filters
directional filters non-directional

tip: two ways of stretching - for display only (no change in DNs), and in file (DNs change) 20

Image enhancement - stretching

Raw image Linear stretch

21

Image enhancement - stretching

Raw image Histogram equalization stretch

22

Image enhancement - filtering

Filtering - how does it work? For each pixel in the image

DN values are multiplied by corresponding values of the superimposed kernel these values are summed the sum is divided by the sum of kernel values

23

Image enhancement - filtering

Filtering - what does it do? Low pass kernel - image smoothing

190

72

24

Image enhancement - filtering

Filtering - what does it do? High pass kernel - image sharpening - e.g. edge enhancement

25

Image enhancement - filtering

Filtering - what does it do? High pass kernel - edge detection

-1 -1 -1 -1 8 -1 -1 -1 -1

26

Image enhancement - filtering

Filtering - what does it do? Directional filters - directional edge detection

-1 2 -1 -1 2 -1 -1 2 -1

27

Image enhancement - filtering

Filtering - what does it do? Directional filters - directional edge detection

-1 -1 -1 2 2 2 -1 -1 -1

28

Single image classification: density slicing

Density slicing - what is it?

DNs of a single band are divided into a number of user-defined intervals or slices. the number of slices and boundaries between them depend on type of land covers of the area. obtained map is displayed as thematic map (i.e. pixels are no longer characterized by DNs, but by class names, e.g. water, snow, dense vegetation)

29

Single image classification: density slicing

Density slicing - what does it do?

Basic way of classifying single band images only classes that strongly differ in spectral response

three classes: water, vegetation, bare soil 30