Coherent pixel selection using a dual-channel 1-D CNN for time series InSAR analysis

Abstract

Coherent pixel (CP) selection is an important step in the processing chain of time series InSAR analysis. In this research, we propose a light deep learning framework, i.e., a dual-channel one-Dimensional Convolution Neural Network (1-D CNN) to select CPs. The 1-D CNN has simple input: SAR amplitude and interferogram coherence, and can be trained with CP samples generated by traditional thresholding method. In an experiment based on Sentinel-1 temporal images in Tianjin, China, the 1-D CNN substantially outperforms the thresholding method and the StaMPS method in terms of the amount and the quality of selected CPs. Additionally, a new measure is proposed to quantify CP quality, which is very useful when other reference data is unavailable. The proposed 1-D CNN framework on CP selection is reliable and fast, and of great significance in developing automatic time-series InSAR processing system.

Keywords

• Coherent pixel selection
• Deep learning
• One-dimensional convolution neural network (1-D CNN)
• Time series InSAR