Digital Image Data Format

BSQ, BIL, BIP, and run-length encoding

1. Band sequential format (BSQ)
   a single band covering the entire scene is stored as one file
   convenient when only selected bands are needed

   TM and SPOT use BSQ format

2. Band interleaved by line (BIL)
   Lines at the same position on individual bands are stored together
        - line 1, band1; line 1, band2; line 1, band3,; etc.
           useful when all bands are needed otherwise can be inefficient

3. Band interleaved by pixel (BIP)
   Pixels at the same position of individual bands are stored together
        - pixel (1,1), band1; pixel (1,1), band2; pixel (1,1), band3
           useful when all bands are needed

4. Run-length encoding
   number of cells, digital number
   a standard form of data compressing algorithm on UNIX and net
   inefficient when data are extremely heterogeneous
 
 
 

Image Rectification and Restoration

1. Geometric correction
        For raw image rectification and register multi-dates or
        multi-spatial resolutions images or data layers

       systematic distortions vs. random distortions
 

Ground Control Points (GCP)
        - features with known locations on a map (X, Y coord)
        - the same features can be accurately located on the image as well (row, col)
        - the features must be well distributed on the map and the image
        - e.g., highway intersections, corners of dammed lakes, etc.

       Coordinate transformation equations
        - relates geometrically correct map coordinates to the
            distorted image coordinates by least-squares regression
          x = a₀ + a₁X + a₂Y;   y = b₀ + b₁X + b₂Y.

       Root Mean Square error (RMS) = Ö(dx)² + (dy)²

Resampling
        - matches the correct output matrix to the distorted image
        - DN of a pixel in the output matrix is based on the DN of
            its surrounding pixels in the distorted image

       Nearest neighbor resampling
        - the DN of a pixel in the output matrix is assigned as the
            DN of the closest pixel in the distorted image
        - advantage:
                simple computation
                maintain the original DN values
        - disad: spatial offset up to 1/2 pixel

       Bilinear interpolation
        - distance-weighted average of the closest 4 pixel DNs
        - ad: smoother output image than the nearest neighbor
        - disad: alter the original DNs

       Cubic convolution resampling
        - uses 16 DNs of the closest pixels, adjusted by distance
        - ad: smoother than the nearest neighb. sharper than bilinear
        - disad: alter the original DN values

       rectification before vs. after image classification
 

2. Radiometric correction
   Radiometric responses differ by dates and sensor types
   correction is necessary when using multi-images

   Sun elevation correction:
                        DN
        -------------------------, assuming the terrain is flat
        Sin(sun elevation angle)
 

   Earth-sun distance correction
        Irrandiance decreases as the square of the earth-sun distance
                E₀ cosq₀                          E - normalized solar irrandiance
        E = -----------,                        E₀ - solar irrad. at mean E-S dist.
                        d²                            q₀ - sun angle from the zenith
                                                        d - E-S dist. in astronomical unit
 

   Atmospheric correction
                  rET                               r - reflectance of target
        Ltot = ----- + Lp                     E - irrandiance on the target
                    p                                  T - transmission of atmosphere
                                                        Lp - scattered path radiation
 

   Haze compensation
        The DN value of an object with 0 reflectance = Lp
        subtract the DN from the band
 

   Conversion of DNs to absolute radiance values
        Necessary when compare different sensors, or relate ground
        truth to image data
        L = (L_MAX - L_MIN)/255 * DN + L_MIN
 

3. Geometric restoration
       Stripping
        - use histogram to identify the defective detector
        - use gray scale adjustment to correct the strips

       Line-drop
        - using average of above/below lines

       Bit errors
        - 3x3 or 5x5 moving average
 

4. Reading: Chpt 7
 

...Back to Ling Bian top page.