Spatial contextual Gaussian process learning for remote-sensing image classification

Abstract

Compared to state-of-the-art classifiers, the Gaussian process classifier (GPC) offers several attractive properties such as the possibility to estimate the hyperparameters or to learn the best input features in a fully automatic way. However, till now, the integration of spatial contextual information in a GPC model for classifying remote sensing imagery has not yet received a sufficient attention compared to other classification approaches for which it has been shown that the classification accuracy can benefit from the exploitation of such information. In this context, in order to improve the GPC capabilities, we propose here to reformulate the GPC learning model so as to integrate spatial contextual information. All the mathematical developments leading to the proposed spatial GPC (SGPC), which embeds iteratively the spatial contextual information in the classification process, is described. Experimental results show that the SGPC can help in improving the classification accuracy compared to the baseline GPC.

Land Cover Classification With Gaussian Processes Using Spatio-Spectro-Temporal Features

“…This point has not been investigated so far for SITS classification at large scale. Some works have been proposed to use contextual information with GP [58], [59] but spatial dependency was limited to a close neighborhood (e.g., 5×5 pixels) with a small training set size. Third, we report an intensive large scale classification benchmark with conventional methods and recent deep models.…”