Trend Change Point Detection in InSAR Derived Displacement Time Series Using MALkCNN: A Deep Learning Approach

Abstract

Interferometric Synthetic Aperture Radar (InSAR) has become a reliable method for estimating ground surface displacement. Change point detection in time series is one of the key post-processing tools for displacement analysis. Due to the complex characteristics of displacement time series, researchers have used statistical and artificial intelligence-based methods to identify change points in the trend of time series and face many limitations. Regarding the lack of reliable ground-truth, a data simulation scheme for the learning stage in deep learning approaches is a challenging issue that is the first focus of this work. Inspired by the generative adversarial network (GAN), we proposed a workflow to simulate datasets, where the real-world time series component with known change points is used in the generative stage and the multilayer perceptron in the discriminator stage. Additionally, a novel method called Moving Average Large Kernel Convolutional Neural Network (MALkCNN) is proposed for trend change point detection, which has a large kernel size to maintain global trend changes while reducing the effects of fluctuations. The results show that the MALkCNN model achieved an F1 score of 0.8321 and an accuracy of 0.9095, outperforming the Time Gated Long Short-Term Memory (TG-LSTM) model as a benchmark while performing tasks seven times faster. The validity of the proposed method has been checked according to the sorting data from the European Ground Motion Service (EGMS) portal located in Germany and Italy. According to our findings, our approach provides a promising solution for identifying trend change points in complex InSAR time series. © Deutsche Gesellschaft für Photogrammetrie, Fernerkundung und Geoinformation (DGPF) e.V. 2025.