Unsupervised noise-robust feature extraction for aerial image classification

The rich data provided by satellites and unmanned aerial vehicles bring opportunities to directly model aerial image features by extracting their spatial and structural patterns. Although convolutional autoencoders (CAEs) have been attained a remarkable performance in ideal aerial image feature extraction, they are still challenging to extract information from noisy images which are generated from capture and transmission. In this paper, a novel CAE-based noise-robust unsupervised learning method is proposed for extracting high-level features accurately from aerial images and mitigating the effect of noise. Different from conventional CAEs, the proposed method introduces the noise-robust module between the encoder and the decoder. Besides, several pooling layers in CAEs are replaced by convolutional layers with stride=2. The performance of feature extraction is evaluated by the prediction accuracy and the accuracy loss in image classification experiments. A 5-classes aerial optical scene and a 9-classes hyperspectral image (HSI) data set are utilized for optical image and HSI feature extraction, respectively. High-level features extracted from aerial images are utilized for image classification by a linear support vector machine (SVM) classifier. Experimental results indicate that the proposed method improves the classification accuracy for noisy images (Gaussian noise 2D σ=0.1, 3D σ=60) in both optical images (2D 87.5%) and HSIs (3D 85.6%) compared with the traditional CAE (2D 78.6%, 3D 84.2%). The accuracy loss in classification experiments increases with the increment of noise. Compared with the traditional CAE (2D 15.7%, 3D 11.8%), the proposed method shows the lower classification accuracy loss in experiments (2D 0.3%, 3D 6.3%). The proposed unsupervised noise-robust feature extraction method attains desirable classification accuracy in ideal input and enhances the feature extraction capability from noisy input.