Sinkhole detection in InSAR interferograms using CNN and LSTM

Code name: X-BBox Author: Anurag Kulshrestha Purpose: Code for extracting training tiles and using U-Net and CNN-LSTM to learn and classify sinkhole related fringe patterns in wrapped interferograms. Author: Anurag Kulshrestha Date created: 03-07-2022 Last modified: 02-02-2023

Info:

1. The traning datasets are created using the XBBox method defined in function: make_training_tiles.
2. The training samples and labels are stored with file names beginnnig with 'trainX_' and 'train_Y' respectively.
3. The models are trained using TSx spotlight data, and tested on Sentinel-1 data.
4. For interferometric processing of TSx-spotlight data, please see TSx_spotlight.py
5. Functions for reading doris derived datasets are written in the doris_read_data.py file
6. The models are declared in models.py

Abstract of related paper:

Many sinkholes are well characterized by elliptical Gaussian-shaped fringes in wrapped Synthetic Aperture Radar (SAR) interferograms. Detection of these patterns over large sinkhole-prone areas remains challenging, especially due to the unavailability of training datasets. Over the past few years, Con- volution Neural Networks (CNN) have proved to be powerful to learn and detect spatial patterns in images. Similarly, Recurrent Neural Networks (RNN), such as Long Short Term Memory (LSTM), have the capability of learning hidden patterns in multi-temporal sequences. As a synergy, this study proposes the use of spatial modelling with U-Net and spatio-temporal modelling with Convolutional Neural Network-LSTM (CNN-LSTM). We ex- tract training datasets from real SAR interferograms created using X-band TerraSAR-X spotlight SAR datasets of resolution 0.23×0.94 m and augment the data in scale-space using a novel method which we call Extract using Bounding Boxes (XBBox). Using transfer learning, we test our trained models on real C-band Sentinel-1 datasets of 20 × 4 m resolution. This was done over a study site near Wink, Texas, USA, where large subsidence was recorded around a sinkhole of ∼500 m diameter in 2015. We used 12 TerraSAR-X and 15 Sentinel-1 SAR images separately acquired between April-2015 and March-2016. The results show that the sinkhole site was detected successfully using U-Net with a weighted average F1-score of 0.98. CNN-LSTM showed relatively lower, but still high accuracy with a weighted average F1-score of 0.92. It was seen that sinkhole detection probability in- creased with the increase of temporal epochs of the input dataset.