|
This final step is performed by the Resampler and it involves both the practical execution of the geometric correction (“re-sampling”) and the creation of daily mosaics on the base of the available (and partly overlapping) L2-registrations (“compositing”). For a given day, the Resampler first creates a new L3 output directory DATA\LEVEL3\ YYYYMMDD , to where it then copies the full contents of the AUXILIARY\ TEMPLATE-directory, i.e. the files IMAGE, GEOG, ANGLES and LEADER. (§ 2.5.2). GEOG is the template grid λβ[CRo], which needs no further modifications. ANGLES and IMAGE are still “empty” (the latter with the exception of the land/sea bit 16 of band 1) but they are already configured in the right way and their values will now be filled with those of the L2-registrations available for the concerned day.
The L2-images are then considered one by one, and for each of them the Resampler first combines the (updated) L2 GEO_GRID CRi[λβ] with the template grid λβ[CRo] and so it creates a new grid CRi[CRo], which retrieves for each template (output) pixel the corresponding position in the L2 input image. To avoid the need for scanning the entire template, the Resampler delineates the limits of the L2-area on the template, and only the L3-pixels in this overlapping zone will be inspected. For every L3-pixel in this zone, the corresponding L2-values (5 bands + mask data) are read via the grid CRi[CRo]. For this “re-sampling”, the user has the choice between two methods: “nearest neighbour” (no interpolation) or “cubic convolution” (interpolation between the values of the surrounding 3x3=9 L2-pixels). The choice only holds for the “real” data, the categorical mask values are always re-sampled with nearest neighbour. In practice, for the MCYFS the nearest neighbour methods is always used.
Most often the concerned template pixel (CRo) can be “filled” with different sets of values, coming from the different L2-registrations of the concerned day. The goal of the “compositing” is to select the “best choice”. SpacePC uses the following compositing criterion (1):
(4-12)
with σ the absolute value of the pixel scan angle (approximate values stored in the L2/L3-masks), NDVI the Normalised Difference Vegetation Index (equation 2-1), and the subscripts 2/3 referring to the L2/L3 image pixels (L3: the values already in the template, re-sampled from a previously inspected L2). If the above condition is fulfilled, then all the L2-information of the pixel is copied to the L3, i.e. the 5 band-values and all mask data with exception of the land/sea bit derived from the classification (the SURFACE_COVERAGE land/sea-mask of the template is kept in the L3). Moreover, the sequential ID-number of the concerned L2-registration (L2-ID) is coded in the band 3 mask bits of the L3. In this way each pixel's origin can later be traced in the L3 final product (L3-pixels with L2-ID=0 were covered by none of the available L2-images).
Once the L3-IMAGE has been established in this way, the L3 ANGLES-files is updated as well. For each L3 ANGLES pixel (a sub-sample of IMAGE), the following variables are respectively searched: the corresponding L3-IMAGE pixel, the ID-number of the L2-image, the corresponding L2-pixel, and the angular information of that pixel in the L2 ANGLES file. The obtained values ( θs, φs, θv, φv) are then stored in the L3 ANGLES.
To conclude: at this level SpacePC follows a quite remarkable strategy. All other comparable AVHRR software packages first create fully corrected (also geo-registered) products for each individual registration, and then apply the compositing (creation of daily S1-mosaics) with the Maximum NDVI criterion (in case of equal choice often slightly conditioned by a minimisation of the scan angle). SpacePC however combines these two crucial steps (geometric correction and compositing) in one single operation. This approach certainly optimises the operation speed and disk requirements, but it has the drawback that the evaluation of the performance and quality of both operations (especially of the geometric correction) is nearly made impossible. In addition, SpacePC's compositing rule (equation 4-12) is merely a Minimum Scan Angle criterion, which certainly lacks the merits of the Maximum NDVI technique (selection of cloud free pixels, whatever the scan angle).
(1) Unfortunately, the condition “IF(L3 is Cloudy)” was eliminated in the code (placed in comment) by EOS in 1998.
|
