Apport des données ASTER et d'un réseau de neurones à rétropropagation à la modélisation de la dégradation du sol d'un bassin marneux du Rif marocain

Abstract

The study of soil degradation phenomena requires the characterization of surface properties. This characterization is based on the integration of the descriptive variables of the local environment: physical, biological and anthropic. In Morocco, the spectacular expansion of erosive processes shows increasingly alarming aspects. Faced with this situation, the country has an urgent need for evaluating the effects of erosion on soil productivity to measure its amplitude and to foster the implementation of better conservation strategies. The overall objective of our research project is to develop a geomatics approach for the mapping and determination of the surface states in a small Mediterranean watershed that can be used for modeling the soil degradation parameters. In the present research, the class definitions of the different levels of degradation are based on the GLASOD ( Global Assessment of Soil Deterioration ) method. To reach the objectives set for our study, we adopted two approaches. The first is a spectral approach and is based on spectral indices, the SAM (Spectral Angle Mapping ) method, and on spectral unmixing. This procedure gave us a better comprehension of the synergistic relationship between the spectral properties of soils and their state of degradation. Also, we elaborated a new spectral index referred to as the LDI (Land Degradation Index ). This index is interesting and provides quantitative results. We were also able to establish an interesting correlation (R[superscript 2] =0.67) between a spectral index derived from ASTER data and the clayey fraction of the soil surface horizon in our study area. The second approach focuses on the modeling of multisource data through the use of a supervised neural network based on a back-propagation algorithm. The quality of results of this approach depends on the determination of the intrinsic parameters of RN. In the present research, we provide some answers concerning issues relating to these parameters in the hope of better defining them. The comparison of the results obtained using the different approaches outlined the existence of a global correspondence between these parameters, in the sense that they represent ground reality. Moreover, the neural approach stands out with a higher accuracy rate (Kappa =0.91) due to the integration of multisource data and spectral information. The Kappa coefficient permitted the evaluation of the accuracy of the classification and its interpretation was carried out based on the calculation of the Z statistics test. The validation of results is based on photointerpretation, ground truth and the analysis of the spatial variability of the data using the geostatistics approach."--Résumé abrégé par UMI.