Words from Afterdark

I have the pleasure to invite you to the defence of my doctoral thesis “Semi-automatic rural land cover classification from high-resolution remote sensing images“, which will be held on Friday, March 24th 2006, at 14:00 in the Genot hall of the Institut Géographique National, in Saint-Mandé, near Paris, France.

The rest of this entry contains a short description of this thesis, the defence site’s address, maps to get there, the defence committee member list, and the long abstract.

This thesis presents a complete image analysis system which, from high-resolution 3 or 4-channel digital images (50cm, colour and optionally near infrared), and using the cadastre database, segments the images into agriculturally-homogeneous regions (fields, forests, vines, and so on), and classifies these regions, tagging each classified region with a confidence measure which indicates the system’s confidence in each classification. It includes a study of the value of texture features and transformed colour spaces for segmentation and classification, two methods for registering a graph onto an image, a novel probability model and associated per-region classification algorithms, and a high-precision period and orientation estimator.

The site’s address is: Institut Géographique National; bâtiment N, salle Genot; 2-4 avenue Pasteur; 94165 Saint-Mandé; France. Please click on the thumbnails to get maps to the IGN.

The defence committee’s members are Dr. Matthieu Cord (ENSEA), Pr. Josef Kittler (University of Surrey), Dr. Jean Louchet (INRIA), Pr. Ferran Marqués Acosta (Universitat Politècnica de Catalunya), Dr. Marc Pierrot-Deseilligny (Institut Géographique National), and Pr. Georges Stamon (Université de Paris 5).

Long abstract:
Much research has been done on the subject of automatically determining land use and land cover in rural areas from images. However, error rates have always been too high for industrial application.

The French National Mapping Agency (Institut Géographique National, IGN), is interested in automatic land cover classification for the purpose of speeding up production of high-resolution topographic maps. The context at IGN is, however, slightly different from that at most other research into automatic land cover classification: First, very high spatial resolution digital images are available (50cm per pixel), but these images have a low spectral resolution (red, green, blue, and in some cases near-infrared channels), which precludes the use of standard hyperspectral classification techniques. Second, cadastre data is available; this data gives a very rough indication of field position. Finally, IGN’s interest stems from practical goals: to reduce the amount of time spent by human photo-interpreters doing manual classifications. Therefore, IGN is not especially interested in obtaining a medium-quality land cover classification of the whole region of interest but would rather have a very high quality classification of only a portion of the terrain; the first one would have to be double-checked in full by a photo-interpreter, whereas the second one would not need further human verification, so that photo-interpreters would be able to concentrate on the remaining areas.

In this Ph.D. thesis, I present a complete image analysis system which, from high-resolution 3 or 4-channel digital images (50cm, colour and optionally near infrared), and using the cadastre database, segments the images into agriculturally-homogeneous regions (fields, forests, vines, and so on), and classifies these regions, labelling each classified region with a confidence measure which indicates the system’s confidence in each classification, and which can be used to filter out regions that are more likely to have been incorrectly classified.

The process starts with a hierarchical segmentation, using a colour space, texture parameters, and shape criteria adapted to the problem of segmenting agricultural regions. This segmentation is used to register the cadastre onto the image, giving large, usually homogeneous regions. Through this registration, the system can also be used to update older classifications. Then, each of these registered cadastre regions —or, if cadastre data is not available, small regions obtained by watershed segmentation— is classified using novel probabilistic per-region classification algorithms which, unlike traditional per-pixel algorithms, do not produce salt-and-pepper noise, and which also output a classification confidence measure for each classified region. These classification algorithms are supervised, and need to be trained beforehand with a ground truth defined by a photo-interpreter.

As an end product we get an image segmentation into classified agriculturally-homogeneous regions, and confidence measures for each part of the segmentation, which a human photo-interpreter can use to correct these results or to concentrate limited available time into the most likely errors.