Context and objectives
This research project aims to improve techniques of remote sensing image processing for land cover mapping and their integration into a geographic information system. The quality and the updating of these inventories are to be the basis of a sustainable and efficient land management.
The purpose is to develop a method improving the recognition and the interpretation of land cover, in order to improve the use of spatial information and ancillary data during the interpretation of multi-sensors and multi-sources data, to preserve classification rules, and to stock the expertise resulting from the interpretation in order to use it again (updating or more detailed inventory on a larger scale).
Two kinds of numerical classification are developed :
During this first phase, classifications by pixel are computed. The aim was to show the importance of the integration of textural and contextual information during the classification. In the same way, the artefacts created by the use of this kind of spatial information are suppressed thanks to the use of a multiple classifier. In spite of the introduction of the texture and the context as well as the use of a complex classifier, the « salt-and-pepper » effect persists. Some post - classification filters will be used to generalise the picture, to increase the level of abstraction of this classification and so to come closer of the visual interpretation CORINE Land Cover.
During the second phase (see ARIOS – phase 1 - T4/11/41), the accuracy of the classification were improved with the use of objects. The images are segmented in regions while using techniques of borders detection. New features of texture and context were computed.
When regions are obtained and the selected discriminative features computed, the objects are classified on basis of different kind of classifiers. The degree of abstraction and the level of generalisation will come closer of the interpretation visual.
Expected scientific results
The first group of techniques tested gives the following results : the advantage to use a tree decision classifier is that it can manage a lot of attributes (this is a necessity because of the relevant size of the window vary according to the land cover class) and of several types at the same time (nominal, ordinal and cardinal). The classification based on spectral attributes obtained with the tree decision give almost similar accuracy that the maximum likelihood analysis. The addition of textural and contextual variables improves very considerably the accuracy. For the tree decision classification, the kappa accuracy coefficient is 0,65 for a few textural variables, 0,72 for a lot textural variables and 0,76 with contextual variables. This result is improved (K = 0,84) when the problematic classes, as “dump sites”, “ transitional woodland-shrub ”, “Sport and leisure facilities”, have been removed.
The second group of techniques (combination of classifier) leads to a very high accuracy, higher than the one of each of the classifiers taken alone (K = 0,88), but the classified image shows very strong artefacts at the objects boundaries. The problem is that the definition of training and validation areas are too “pure”; therefore the system does not learn to recognise these boundaries and they are not taken into account for the validation (the accuracy is therefore overestimated). To solve this problem, training and validation areas have been choosing at random within the interpreted images derived from the CORINE Land Cover database. The classified image is much closer to CORINE Land Cover data even though the kappa coefficient is a bit lower (K = 0,77)