Scene classification in remote sensing images using a two-stage neural network ensemble model

Scene classification has long been a challenging task in the remote sensing field. Conventional approaches based on hand-crafted features are not suitable in large scale remote sensing images. Convolutional Neural Network (CNN) achieves great success in computer vision field by learning hierarchical features automatically from mass data. However, the shortage of labeled dataset in remote sensing field results in severe overfitting and the ensemble of several networks have better generalization ability than one single network. In this letter, we propose a novel Two-Stage Neural Network Ensemble Model to solve the problems mentioned above. Firstly, to overcome overfitting, we pre-train a CNN using the ImageNet dataset and fine tune the network by labeled remote sensing images. Then, the output features are fed to a Restricted Boltzmann Machine (RBM) Retrained Network to get better feature representations. Finally, in testing stage, a method based on Ensemble Inference Network (EIN) is introduced to enhance the generalization ability by combining the classification results of several networks. Experimental results on the UC Merced Land Use (UCML) dataset demonstrate the effectiveness of our proposed method.