Evaluation and predictability of soil susceptibility to water erosion at the regional scale using hyperspectral data and external knowledge

International audience The problem of environmental risk management is becoming a great concern worldwide. Identification of vulnerable areas and selection of prevention policies can minimize damages systematically, efficiently and more economically. In this paper we propose a new hybrid approach based on two main stages that combines spectral images unmixing and information fusion for natural risk prediction. First, a method using the Spectral Angle Mapping (SAM) technique for abundance quantification may be considered for hyperspectral remote sensing images unmixing. Then, we outline a neuro-fuzzy data fusion strategy based on the synergy of artificial neural networks and fuzzy inference systems. This methodology has been proposed to deal efficiently with the highly subjective nature, the imprecise and the non-linearity of data involved in this serious problem. The experimental results show an application example of the proposed methodology using a Hyperion image and external data. A case study was conducted on a subset image located in southern Tunisia regarding the "Hydrous Erosion " phenomena. A data set collected from 45 sampling sites in the study area has been used to train and validate the hybrid model. We compare our results with the ground truths by the evaluation of the classification accuracy using fuzzy confusion matrix. Up to 91.9% of sampling sites have been correctly predicted with neural-fuzzy model. Such performance proved to be very competitive when compared with other approaches such as fuzzy logic prediction and neural networks, the predictive capacities were 75.04% and 82.01%, respectively. The proposed methodology can support decision-makers in evaluating and risk vulnerability prediction.