The evolution study of the natural hazard zones by the Kohonen self-organising map classification

This article consists to explain the application of a semi automatic and multi temporal method of classification of the spot images. The objective is to follow the evolution of the units of landscapes factors of risk in the Nouakchott area (in West Africa). In fact, it was proved in the literature of image processing techniques and remote sensing, that there is not a completely automatic method of classification, considering the complexity of the phenomena represented by a pixel and the geometrical resolution of this same pixel. The method used in this work allows the intervention of the expert (of validation) in certain stages of the application. The goal is to improve the results of classification without repeating all the calculation. We applied the method to chart the landscapes units of the environment of Nouakchott. The objective is to monitor the environmental hazards in this region characterizing the classes corresponding to landscape units present in SPOT images and to survey their temporal evolution. The method consists to apply the Kohonen Self-Organizing Map (SOM) algorithm on a data base of multi-dates images. The obtained Kohonen topological map is constituted of 100 neurons. Each neuron represents the reflectance spectrum of a subset of the training database. This map is thus representative of all the units of landscape present in the images of various dates. A Hierarchic Ascending Classification (HAC), with number of classes, is applied on the topological map to aggregate the subsets presenting similar reflectance, therefore to group all the local variability of the landscape. The classes gathering the 100 calculated subsets are labelled to make the correspondence of a given class to a landscape unit on the ground. These units present the factors element of risk. Once the map is validated, for each date, the image is classified automatically by a simple assignment of its pixels to the nearest neuron of the topological map. We thus obtain a series of classified images from the same map, what allows one to quantify the evolution of each class compared to its counterpart of other dates, and therefore to evaluate the risks evolution The classes gathering the 100 calculated subsets are labelled to make the correspondence of a given class to a landscape unit on the ground. These units present the factors element of risk. Once the map is validated, for each date, the image is classified automatically by a simple assignment of its pixels to the nearest neuron of the topological map. We thus obtain a series of classified images from the same map, what allows one to quantify the evolution of each class compared to its counterpart of other dates, and therefore to evaluate the risks evolution