Now that all SPOT4(Take5) images have been processed (pfew !), we can make an appraisal of the performances. Let's start by the geometry, which caused us a lot of trouble :
- SPOT4 has a location accuracy around 400 mètres, but during the experiment, it went through a fifteen day period when the location errors could reach 1500 m.
- We seek a multi-temporal registration performance of 0.3 pixel RMS. This performance is difficult to measure because the measurement technique itself (correlation image matching) is not perfectly accurate.
- We provide as a criterion the maximum registration error observed for the 50% best results or for the 80% best results. It is likely that the last criterion includes less inaccurate measurements.
Here are the observed performances for 3 very different sites :
- CMaroc site, which is an arid site with a green period in march, a lot of blue sky, and high mountains (the Atlas). performances are excellent, with errors lower than 0.3 pixels for 50% of the measurementsl.
- CBretagneLoireE site, which is a rather flat coastal area with large tides, and is often very cloudy. In that case, performances are still better than 0.5 pixels. The worse dates correspond to images with a large cloud cover, for which it is not easy to automatically collect accurate ground control points.
- JSumatra site is a very flat area, covered with very uniform equatorial forest, and a large river whose limits change with time. In that case, the performance is really bad, with registration errors up to 10 pixels. This uniform site does not enable to find good control points, and the ones that are found are often along the river whose contour changes with the water level.
We have obtained very good results for most sites, with registration errors below 0.5 pixels (10m) even when the initial location error reaches 1500m. However, 4 sites are resisting to this processing. These 4 sites correspond to flat forest sites covered by equatorial forest : JSumatra, JBorneo, EGabon, ECongo. The ECongo site is even so uniform that it is not possible to measure its registration performance.
These sites will be distributed with the others in a few days with the first version of the products, but you should use them cautiously.
Finally, if the registration of 95% of images is good, the location performance is inherited from our reference images, ie LANDSAT (5 et 7). The next versions will be based on Geosud (IGN) images in France and on LANDSAT 8 data elsewhere. Performances should be enhanced in the next versions.