Information Extraction and Interpretation for multi-scale Remotely Sensed Imagery – XIMRI

Remote sensing is a key technology for the observation of the Earth. The interpretation of remote sensing images is a challenging task. In the past
decades, the resolutions of remote sensing images increased drastically. From MODIS (200m), LANDSAT (30m) to QuickBird (0.6m) and GeoEye (0.5m),
the spatial resolutions of the remote sensing images increased as the high/very high resolution satellites came into operation. These increased resolutions provide an increased amount of information and enable a very accurate analysis and interpretation of the sensed scene. However, the increase of the resolution requires the development of new processing algorithms, as most of the standard methods used with low to medium resolution data fail when confronted to very high resolution images.

In order to address successfully all these new challenging applications, three major problems emerge:
On one hand, the radiometric features classically used for characterizing the remote sensing images are not adaptive to the high/very high resolution images. The contextual features must be used.
On the other hand, most of the analysis method developed one resolution or one generation of instrument can't be directly used to other resolutions: a unified method which is able to classify the images which is independent from the spatial resolution of the image is therefore necessary.
Finally, in order to classify the remote sensing images with the help of different types of features (radiometric and contextual features), the way for the fusion of these features must be inspected.

Based on the these issues addressed above, we propose a research project entitled: Information Extraction and Interpretation for Multi-scale
Remotely Sensed Imagery (XIMRI)

At fisrt (WP1), we have to extract the features from the image. Since the radiometric features are classical, we are more interested in contextual
features in this project.
Secondly (WP2), we have to compute the features which are indepedent to the spatial resolution of the images. For the case of contextual features extracted from high/very high resolution images, we will propose the multi-scale features which characterize the geometry of the structures of different scales. For the radiometric features, we will explore the pansharpening technology which allows the fusion of the spectral information of the images at different resolutions. For the hyperspecral images, the spectral unmixing which extracts the spectral information potentially invariant from the resolution change (for the linear mixture case).
For the last step which is classification, we will propose a machine learning approach. The fusion of different features will be implemented by a combination of kernles. How the kernels adapted to different features are combined is the objective of WP3.