• Automatic co-registration of repeated acquisitions of the same area
• Automatic change detection in cloud-contaminated optical satellite images

Automatic co-registration

NR has developed a method for automatic co-registration (Eikvil et al., 2009) of images from repeated acquisitions of the same area on the ground. This method will be tested for images of tropical forest areas. To begin with, we will focus on tropical rain forest areas in Brazil. It is important to understand under what circumstances the method the method will fail. Obviously, if an acquisition has (close to) 100% cloud cover, there is no way to co-register that image to a reference image in a meaningful way. On the other hand, we know that the method can handle a large fraction of cloud cover. Further, if the scene has mostly undisturbed forest, it might be difficult for the method to identify suitable landmarks in the images.

Automatic change detection in cloud-contaminated images

NR has developed a method for vegetation mapping based on optical images with missing data (Salberg, 2009). The missing data could be due to clouds, cloud shadows, snow, and Landsat 7 SLC-off sensor failure. NR is also developing a landcover change mapping method based on Markov chains. These methods will be combined and adjusted to tropical forest change detection. Central to the methods is the availability of adequate training data, in the form of in-situ measurements and/or very high resolution optical satellite images. Since automatic land cover classification will make some mistakes, change detection should not be based on simply computing the difference between two land cover classifications form different dates. Rather, a moderately long time series should be used to detect forest disturbances as deviations from changes that occur naturally, due to growing season cycles, changing weather, sun elevation, etc.