Mapping the Baltic Sea Basin

1

• Mapping the Baltic Sea Basin

By Michael Ledwith, Lantmäteriet Metria
Miljöanalys michael.ledwith_at_lm.se tel 46 8 579
972 98
Countries entirely covered Belarus, Czech
Republic, Denmark, Estonia, Finland, Germany,
Latvia, Lithuania, Norway, Poland, Slovakia,
Sweden Countries partially covered Austria,
Belgium, France, Hungary, Luxembourg, Moldova,
Netherlands, Romania, Russia, Switzerland, Ukraine
2
SPOT Vegetation Images

• Classification of the Baltic Sea Basin was
  carried out using low resolution (2250 km swath
  width) images acquired by the Vegetation
  instrument on-board the SPOT satellite.
• Four spectral bands
• Blue 0.43 - 0.47 ?m
• Red 0.61 - 0.68 ?m
• Near Infrared (NIR) 0.78 - 0.89 ?m
• Mid Infrared (MIR) 1.58 - 1.75 ?m

• Additional information (based on ground
  reflectance) accompanying the radiometric data
  includes
• NDVI / Status Map / Viewing Zenith Angle / Solar
  Zenith Angle / Viewing Azimuth Angle / Solar
  Azimuth Angle / Synthesis Time Grid

3
SPOT VGT S-1 and S-10 Composites

• North and south of 32º, the VGT sensor acquires
  more than one image per day due to its wide
  swath. The daily synthesis composites (S-1) are
  computed from the different passes of one day
  for each location. In the high latitudes, each
  pass reflects different viewing and solar angles.
  The ten-day synthesis is computed from all of the
  passes for each location acquired during the
  ten-day period. The synthesis is created using
  the following criteria

the pixel does not correspond to a blind or
interpreted pixel, the pixel is not flagged as
cloudy in the status map, the pixel has the
highest TOA NDVI compared with other pixels at
the same location.
4
Pre-processing SPOT VGT data

• conversion of S-10 file format from HDF to
  generic binary,
• importing band data into image processing
  software (e.g. Erdas),
• adding projection information (Platt-Carre) to
  images,
• creating a five-band stacked image of the four
  radiometric bands and NDVI,
• masking out the blind and aberrant MIR pixels,
• masking out the compositing shadow effect at
  the land borders,
• masking out clouds,
• masking out water,
• masking out snow and ice.

5
Blind and aberrant MIR pixels

• The MIR sensor consists of 6 bricks of 300
  detectors in line. At every brick junction there
  is a blind detector. Therefore, inbetween the 6
  bricks there are 5 blind spots.
• The MIR detectors are sensitive to proton
  fluxes coming from the sun. When a proton hits a
  detector, it is disturbed or destroyed, depending
  on the power of the proton or the number of the
  previous chocks it has received. On average, one
  detector is blind or aberrant per month.
• Prior to 17 May 2001 when CTIV initiated a more
  rigorous algorithm to remove these pixels,
  several northeast-southwest stripes are present
  on all of the images.

6
Removing the blind and aberrant MIR pixels

• A simple edge detection filter oriented
  northeast-southwest was applied to the MIR band.
  A minimum (black) and maximum (white) threshold
  was then applied to the image. This was very
  effective in removing the unwanted noise from the
  image.

7
Compositing Shadow Effect

• For the calculation of synthesis images, the
  compositing method utilizes selection of maximum
  NDVI as the main rule. Over land, cloudy pixels
  generally have a lower NDVI than pixels where the
  land surface is visible. As a consequence, the
  compositing method tends to reject cloudy pixels
  over land in favor of non-cloudy pixels. However,
  over water where the cloud pixel has a higher
  NDVI value the situation is reversed and the
  cloud pixel is chosen over the water pixel.
• A land-water mask is applied to the synthesis
  images which is slightly over-dimensioned for the
  land masses in order to cope with local
  inaccuracies. As a result, a roughly four pixel
  wide border of water/cloud pixels surrounds the
  land masses. Additionally, all water bodies (i.e.
  lakes, large rivers) show this effect as well.

8
Removing the compositing shadow effect

• Simple masking of the compositing shadows based
  on spectral signatures could not be applied to
  the image due to the presence of similarly valued
  pixels over the mountainous regions of Norway,
  Sweden and southern Germany (snow, ice, rock
  and/or bare soil).
• Several techniques were used with varying degrees
  of success to remove the effect including
• unsupervised classification using 100 classes
  in order to find a shadow class,
• area specific region growing according to
  spectral properties using 8-directional
  neighborhoods and a Euclidean distance of between
  20 and 250,
• The final mask was constructed by clipping out 75
  small areas along cloud free coastlines of S-1
  images (2000-06-07 and 2000-09-12 to -21). In
  three cases, S-10 images from 2000-06-01,
  2000-06-11 and 2000-09-21 were used.
• These areas were processed using unsupervised
  classification (20 classes). The classes
  corresponding to compositing shadow were selected
  and joined in a composite mask. Clump
  (8-directions) and sieve processes were performed
  to remove snow/ice clusters and extraneous
  pixels. The final map used a sieve of 50 pixels
  initially and then a final sieve of 2 pixels.

Before
After
9
Removal of highly off-nadir pixels

• The SPOT VGT sensor has a linear array which
  greatly reduces the amount of distortion
  associated with off-nadir pixels - especially
  compared with scanning arrays such as AVHRR.
  However, at angles greater than approximately 50
  distortion begins to effect the pixel size.
• Thus pixels which have been acquired at angles
  equal to or greater than 50 should be removed
  from the image prior to processing and
  classification.
• Using ancillary data which accompanies the
  radiometric data (i.e. viewing zenith angle file)
  a mask was created to remove distorted pixels.

• In most cases the removal of these data did not
  greatly affect the overall usability of the
  image. However, sometimes large swaths of data
  were removed and this greatly affected individual
  classifications.

10
Classification of S-10 images

• Several different techniques were tested in order
  to determine the best method of accurately
  classifying the SPOT VGT S-10 images. For the
  most part, unsupervised classification on single
  S-10 composite images performed an adequate job
  at differentiating between the major types of
  land cover (i.e. forest, herbaceous, water, snow,
  sparse vegetation). Urban pixels are very
  difficult to classify using low resolution images
  and were identified using ancillary data.
• Due to several factors, the mountainous regions
  of Norway and Sweden proved to be an exception.
  Additionally, the make up of the forests of
  northern Sweden, Finland and eastern Russia -
  which contain perhaps thousands of lakes and bogs
  - makes it difficult to consistently classify
  these pixels accurately.

• Supervised classification using Maximum
  Likelihood, Mahalanobis Distance and Minimum
  Distance all failed to adequately classify
  western Scandinavia.
• Principal components analysis showed great
  potential in classification but due to time
  constraints this could not be further
  investigated. However, certain generalizations
  can be made relating to a PCA run on VGT
  radiometric data with 5 PC band output.
• PC1 - clearly differentiates between closed
  needle-leaf forest and other forms of landcover,
• PC2 - ideal band for identifying (and removing)
  water and snow pixels - particularly with respect
  to the water/cloud ring.

11
Unsupervised Classification

• The result of the unsupervised classification was
  a single band image consisting of clusters, or
  groupings, of areas in which the pixels show
  similar spectral signatures in most or all of the
  input bands. The results of an unsupervised
  classification, conducted on an S-10 composite
  image for 2000-09-21, are shown in to the right.
  The image is centered on the Polish city of
  Poznan, visible as a dark purple cluster. The
  large swaths of pink are agricultural lands where
  the crops have been harvested. The dark green
  areas are needle-leaf evergreen stands. The
  Bydgoszcz Canal and Notec River are quite obvious
  as a light green linear feature, running
  east-west near the top of the image. Other
  agricultural areas, grasslands and deciduous
  forests are represented as a light green. The
  black speckles along the right side of the image
  are pixels that were masked out during the
  pre-processing of the image.

Results of unsupervised classification of four
band VGT image with 40 classes as output, and a
maximum of 15 iterations and a convergence
threshold of 0.95.
12
Classification Algorithm

• Essentially a matrix is created that cross
  tabulates the data in the reference image with
  the results from the unsupervised classification.
  From this matrix, data such as the majority pixel
  (the most commonly occurring pixel) and the
  percentage of pixels (fraction of majority)
  within a particular cluster can be easily
  obtained.
• During the classification process, a pixel can
  match the criteria for labeling a pixel within
  several of the classification steps. That is, a
  pixel can be assigned to different land cover
  classes. However, all information produced in the
  different steps are merged together into a single
  land cover dataset in a predefined priority order
  (parallelpiping), thus the pixel is assigned to
  the land cover class to which it has the highest
  likelihood of belonging.

13
Comparing Majority and Percentage Cluster Images
1 89 2 7 3 1
1 Croplands 2 Pastures 3 Wetlands
1 97 2 2 3 0
1 Croplands 2 Pastures 3 N/A
Within the majority image, a cluster in the first
band contains a value corresponding to the most
frequent class as compared to the reference data.
Within the percentage image, a cluster in the
first band contains a value corresponding to the
percentage of the cluster that is occupied by the
class as identified in the majority image. In MAJ
1 above, the cluster corresponds to 97 croplands
per the reference data, while the MAJ2 cluster
corresponds to 89 croplands (as well as 7
pastures).
14
Algorithm Results

• A three-band majority image for 2000-09-21. Dark
  green areas are needle-leaf evergreen forests
  while lighter green pixels show agricultural
  areas and deciduous broadleaf trees (Helsinki,
  Finland).

A three-band percentage image for 2000-09-21.
Light areas have a higher degree of agreement
with the reference data (Helsinki, Finland).
15
Correctly Classified Areas
The image to the right shows the correctly
classified pixels in the 2000-06-01 S-10
composite image centered over Copenhagen,
Denmark. Gold pixels are agricultural lands while
the dark green pixels are needle-leaf evergreen
forests. The pixels that have been declared as
correctly classified are removed from the
satellite scenes (masked to a zero value) and a
second unsupervised classification is performed.
16
Using PC Analyses

• The image to the left shows the third principal
  component from an analysis done on the four
  radiometric bands of a S-10 composite image from
  2000-09-21. This component is particularly well
  suited to differentiate between vegetation
  (light) and non-vegetation (dark). In addition,
  snow covered areas are visible as bright white
  regions - due to sensor saturation.

17
Final Correctly Classified Pixel Image

• Here is a mosaic of all the correctly
  classified pixels for 2000-09-21 as acquired
  during the second stage of classification. Gold
  indicates agricultural areas, dark green
  represents needle-leaf evergreen and mixed
  forests, red areas are sparsely vegetated and
  light brown pixels are pastures.

18
Filling in the gaps

• The above left image shows the open evergreen and
  mixed forests. The lower left image shows the
  sparsely vegetated areas, bare rock and snow
  covered areas.
• The above right image shows the wetlands and
  lichen covered areas. The below right image shows
  the mosiac class of mixed forest and croplands.
• For the majority of these images, the individual
  cluster was individually compared with the
  reference data in order to classify the cluster.

19
Assembling the final image

• The final image was created piecewise by
  mosaicking the least specific (or important)
  information first and then adding layer upon
  layer of more specific information. This was done
  to assure that the classes that were more
  difficult to identify and thus acquired using
  different and more specific techniques wouldnt
  become lost within the grosser classifications,
  e.g. pasture being included in the agriculture
  class. It also allowed urban pixels to be placed
  correctly even if an urban area was classified as
  forest (due to the presence of many trees and
  parks).