Performance study of the smart networks for remote sensing image textures identification

摘要

In this paper the performance evaluation of smart networks to identify highly heterogeneous textures remote sensing images was investigated. These networks are Feed Forward Neural Networks (FFNN), Quantum Neural Network (QNN) and Support Vector Machine (SVM). This evaluation is performed through an optimization training time and number of parameters of smart networks in the constraint to achieve optimal identification rate of the textures. The study also concerns the influence of the nature of heterogeneous textures on the choice of smart networks parameters to obtain elementary unit of textures. The objective is to study the impact of the textural information on the network design and considering that the samples of textures have a textural complexity due to the textural correlation and the overlapping rates of species in these textures. Textures bases used in this study are taken from different remote sensing images sources: an airborne radar image and an ASTER satellite whose resolutions are totally different. We have studied the influence of the spatial resolution on the textures identification and network performance relative to each of the two types of images.