Multi-resolution change detection

Yesterday I was at the PhD thesis defence of Amandine ROBIN. The subject of the thesis is "Sub-pixel classification and change detection using multi-resolution remote sensing images" and the work has been co-funded by CNES (the French Space Agency) and EADS-Astrium.

The goal of the work was to develop techniques able to combine the high temporal repetitivity (1 day to 10 days) of low resolution (100 m. to 1 km.) sensors together with the high spatial resolution (10 m. to 30 m.) of sensors having longer revisit periods (from 30 days to several months). This means trying to combine SPOT XS images with MeRIS data, or Landsat with SPOT VGT.

In terms of application field this is very interesting, since, for instance, in the case of precision farming, one may want to update a landcover map several times in one season, but due to cloud cover issues it may be difficult to obtain a series of high spatial resolution data with a good temporal sampling. In this case, the approach proposed in the thesis was to use a high resolution geometric reference for the regions of the study area (the agricultural plots) and perform a classification of each region using the low spatial resolution temporal series.

Typically, this means using a SPOT XS image acquired at the begining of the season in order to perform a segmentation (the classification is nearly impossible with one single date, since most of the classes are distinguished by their temporal evolution rather than by their radiometry at a given date). Then, using several MeRIS images acquired during the season, the classification may be performed. It is interesting to note that the final classification has the SPOT resolution (20 m.), but the temporal series has pixels which have a size of 300m. by 300m. (resolution factor of 15 by 15). Amandine has shown that the results are good up to resolution factors of about 40 by 40.

The other application was detecting changes between a low spatial resolution temporal series and a high resolution reference classification. That means that you have a high resolution classification obtained with the method presented above and using low resolution images (even a single acquisition, no need for a long temporal series) you can detect which low resolution pixels have changed. This change alert can be used to trigger a new classification or even to schedule a high spatial resolution acquisition.

Amandine's work is based on a sound theoretical approach and the results can be applied to optical sensors either having the same spectral bands between the high and low resolution data or even having different bands but allowing to compute the same vegetation index (NDVI, Leaf Area Index, etc.).

I am persuaded that these results are ready to be used in operational chains.

Amandine's homepage has some material available for download which is worth looking at.