Classification of Hybrid Multiscaled Remote Sensing Scene Using Pretrained Convolutional Neural Networks

A novel hybrid multiscaled Remote Sensed (RS) image classification method based on spatial-spectral feature extraction using pretrained neural network approach is proposed in this paper. The spectral and spatial features like colour, texture and edge of RS images are extracted by using nonlinear spectral unmixing which is further scaled by using bilinear interpolation. The same RS image in parallel path is first scaled using bilinear interpolation and further spectrally unmixed. These two paths are further fused together using spatial-spectral fusion to give multiscaled RS image which is further given to a pretrained network for feature extraction. For authentication and discrimination purposes, the proposed approach is evaluated via experiments with five challenging high-resolution remote sensing data sets and two famously used pretrained network (Alexnet/Caffenet). The experimental results provides classification accuracy of about 98% when classified at multiscale level compared to 83% when classified at single scale level using pretrained convolutional networks.