SAR patch categorization using dual tree orientec wavelet transform and stacked autoencoder

Abstract

This paper presents a categorization of Synthetic Aperture Radar (SAR) data patches. The categories of the SAR data were designed manually by cutting several spotlight SAR products into different categories. The supervised approach to the categorization was proposed, where an oriented dual tree wavelet transform was used to decompose energy of the original image. Subbands of wavelet transforms with different orientations were used for computation of spectral features. The log commulants were estimated for each subband and 8 additional rotations were used for feature extraction. Those features were fed into stacked autoencoder (SAE). The SAE was pre-trained by greedy layer-wise training method. Capable of feature expression, SAE makes the fused features more distinguishable. Finally, the model is fine-tuned by a softmax classifier and applied to the categories selection of targets. The proposed method is comparable with the state-of-the art methods for SAR data categorization.