Application of deep learning for coherent pixel selection in time series InSAR for urban area and transport infrastructure monitoring
The application of Time-series Interferometry of Synthetic Aperture Radar (TS-InSAR) for monitoring small and narrow infrastructural elements can be limited when using medium spatial resolution SAR sensors, such as Sentinel-1, due to sparse coverage of coherent pixels. Therefore, carefully selecting...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225003656 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The application of Time-series Interferometry of Synthetic Aperture Radar (TS-InSAR) for monitoring small and narrow infrastructural elements can be limited when using medium spatial resolution SAR sensors, such as Sentinel-1, due to sparse coverage of coherent pixels. Therefore, carefully selecting coherent pixels is crucial to maximise the density of reliable measurement points and to minimise noisy observations. In this study, a novel framework is proposed for selecting coherent pixels within the Stanford Method for Persistent Scatterers Small Baseline (StaMPS-SB) approach based on the temporal coherence predicted by deep learning. Unlike previous complex deep learning approaches to this problem, a multi-layer perceptron (MLP) model is trained on time-domain and frequency-domain features extracted from a time series of SAR amplitude images to predict the temporal coherence value of each individual pixel, and thus to select coherent pixels. A long short-term memory (LSTM) model, trained directly on the same amplitude time series, is used as a benchmark for evaluating the MLP model’s performance. Predictions of temporal coherence from MLP and LSTM models are similar, but the MLP model training time is much faster. Compared to standard amplitude difference dispersion (ADD) analysis, the MLP model increases the number of coherent pixels over buildings, linear infrastructure, and non-urban classes by 40 %, 68 %, and 45 %, respectively. In addition, the MLP model approach produces fewer unreliable (non-coherent) pixels; over non-urban regions, 70 % to 80 % of pixels selected by the ADD analysis have low temporal coherence (<0.8), compared to only 10 % to 12 % selected by the MLP model. |
|---|---|
| ISSN: | 1569-8432 |