Multi-sensor data fusion for supervised land-cover classification through a Bayesian setting coupling multivariate smooth kernel for density estimation and geostatistical techniques

The data fusion is a growing research field, which finds a natural application in the remote sensing, in particular, for performing supervised classifications by means of multi-sensor data. From the theoretical standpoint, to address such an issue, the Bayesian setting provides an elegant and consistent framework. Recently, a methodology has been successfully proposed incorporating a geostatistical non-parametric approach for improving the estimation of the prior probabilities in the scope of the supervised classification. In this respect, a limitation affecting the Bayes computation in the multi-sensor data is the naïve approach, which considers independent all the sensor measurements. Obviously, such hypothesis is unsustainable in practice, because different sensors can provide similar information. Therefore, an enhancement of the previous described method is proposed, introducing the smooth multivariate kernel method in the Bayes framework to furtherly improve the probability estimations. A peculiar advantage of the smooth kernel approach concerns the fact that it is inherently non-parametric and consequently overcomes the multinormality data hypotesis. A case study is presented based on the data coming from the AQUATER project.