Deep Learning-Based Homogeneous Pixel Selection for Multitemporal SAR Interferometry

Abstract

Homogeneous pixel selection (HPS) plays an important role in the application of multitemporal SAR interferometry. The statistical goodness-of-fit testing of the temporal samples has been widely used for HPS. However, the detection rates of the existing methods are unsatisfactory under small datasets. In this paper, a stacked auto-encoder (SAE) network based method is proposed for the selection of homogeneous pixels under the idea of deep learning image classification, as termed by SAEHPS. The SAE network is used to learn the spatial distribution behavior of the average intensity image. The deep network is trained and tested on different high-resolution SAR datasets of the Hong Kong Airport and the Fuzhou City, and three pixel-wise labels (i.e., high, medium, and low reflections) are regarded as outputs of model learning. The unsupervised training and supervised fine-tuning realize the class prediction. The results show that the SAE can achieve robust accuracies above 90% based on empirically labeled samples, especially in non-architectural areas where the distributed scatterers exist. The SAE results are devoted to the multitemporal PS/DS InSAR approach to identify homogeneous pixels. Both qualitative and quantitative experiments in HPS, phase optimization, and deformation monitoring have demonstrated the superiority of the novel method.