Transfer Learning for Multiscaled Classification of Hyperspectral Images using VGG-VDD Pretrained Networks

Abstract

A novel Remote Sensed (RS) image classification method based on multiscaled spatial-spectral feature
extraction using pretrained neural network approach is proposed in this paper. Scaling in spatial
domain is achieved by Laplacian pyramid method and further spectral features are derived from these
scaled images, further these two paths are fused together using spatial-spectral fusion to give
multiscaled RS image which is further given to a pretrained network for feature extraction. This paper
also discusses in detail the use of pretrained networks for classification of RS images instead of building
a complete new Convolutional Neural network (CNN) which requires lot of labeled training data and
a large training time. For authentication and discrimination purposes, the proposed approach is
evaluated via experiments with five challenging high-resolution remote sensing data sets and five
famously used pretrained network (VGG F-M-S/ VGG-VDD-16/ VGG-VDD-19). The experimental
results provides classification accuracy of about 99.041% for VGG-S when classified at multiscale level
compared to 83% when classified at single scale level using transfer learning technique.