Optical satellite imagery is widely used for estimating ground movement in the aftermath of natural disasters such as earthquakes. This type of imagery enables detailed analysis of the factors and mechanisms that drive or influence these events. By using sub-pixel correlation algorithms, it provides precise displacement measurements (in the meter-to-centimeter range) and high spatial resolution (decimeter-to-centimeter level) by comparing images taken before and after the event. In this study, we present a deep neural network approach, trained on our new specific realistic dataset FaultDeform, to retrieve full-scale seismic ground motion displacement fields from optical satellite images with sub-pixel precision. The FaultDeform dataset, available at https://doi.org/10.57745/G02ZXZ, is the first satellite synthetic dataset tailored for ground motion estimation. We introduce the GeoFlowNet pipeline, utilizing a U-net architecture to solve the displacement estimation problem, delivering high-speed performance through GPU implementation, and outperforming current correlators in speed and precision. Comprehensive comparisons with state-of-the-art methods such as COSI-Corr, MicMac and CNN-DIS, and validation on real-world data from the 2019 Ridgecrest and 2013 Balochistan earthquakes showcases the robustness of our method. Codes are freely available: gricad-gitlab.univ-grenoble-alpes.fr/montagtr/GeoFlowNet.

Displacement estimation using remote sensing images is an effective approach for assessing surface displacement caused by natural disasters like earthquakes and landslides. By employing pixel correlation algorithms, high-precision displacement maps can be generated from images taken before and after surface movement. However, traditional methods often rely on spatial regularization or frequency masking to reduce high-frequency noise, which can smooth spatial details and result in biased displacement estimates, especially near sharp discontinuities typical of earthquake surface ruptures. Moreover, sub-pixel displacement estimation using Synthetic Aperture Radar (SAR) images remains a challenge compared to optical images, due to the strong impact of speckle noise. This paper presents GeoFlowNet-SAR, an innovative sub-pixel displacement estimation method leveraging SAR images. SAR offers advantages thanks to an all-weather observation and high penetration, making it suitable for conditions typically challenging for optical systems in the visible light spectrum. This study uses Sentinel-1 SAR Single Look Complex (SLC) images with dual-polarization (VV and VH modes) and Interferometric Wide (IW) swath mode to balance coverage and resolution. By training on simulated displacement datasets with realistic sharp discontinuities, GeoFlowNet-SAR directly predicts surface displacement fields, providing highly efficient, robust, and precise results, while overcoming some limitations of traditional methods. The effectiveness of the proposed methodological contribution is first quantitatively demonstrated using synthetic simulated earthquake datasets, including comparisons with state-of-the-art correlation methods. The method is further validated using two real remote sensing images from the 2019 Ridgecrest earthquake and from the 2023 Turkey-Syria earthquake. The observed results from these real datasets confirm the effectiveness of GeoFlowNet-SAR in practical applications. The codes are available at: https://gricad-gitlab.univ-grenoble-alpes.fr/giffards/geoflownet-sar.

Estimating the ground displacement from non-rigid registration of two optical satellite images, separated from hours to months, is key in the study of natural disasters such as earthquakes. Compared to standard image registration and flow estimation tasks, a key challenge here lies in resolving very small displacements (typically cm- or m-scale) with sub-pixel accuracy and precision using coarser image resolutions (e.g. 15 m for Landsat-8). Traditional block matching/sliding window methods, employing local windowed correlation techniques, are unable to reduce the effects of long-wavelength noise arising from differences in image lightning, vegetation, or acquisition artifacts. By using both local and global scales, fully convolutional deep learning registration models (U-nets) are potentially able to better resolve ground displacements, less affected my multi-scale noise. Yet, no labelled database exists for ground deformation. Here, we develop a new synthetic database of 50,000 realistic satellite image pairs containing simulated earthquake displacements, along with their ground truth displacement maps, which are used to train state-of-the-art fully convolutional deep learning models (U-net). The inference shows good preliminary results, with a fast computation time (less than 1 second for a 256 × 256 image).

Mountain glaciers play a critical role for mountain ecosystems and society with major concerns related to their future evolution and related water resources. Modeling glacier future evolution allows anticipating climate change impacts and informing policy decisions. It relies on accurate ice thickness estimation at regional scales. This paper proposes a deep learning based approach in a supervised learning framework for ice thickness estimation at a regional scale from surface ice velocity measurements and a digital elevation model. A neural network model built upon a ResNet architecture is proposed based on the trade-off between the model complexity and the prediction efficiency. Promising results are obtained from data including 1400 glaciers in the Swiss Alps, highlighting the potential of deep learning based approach for large scale ice thickness estimation. The incorporation of expert’s knowledge into the neural network model further helps refine the model prediction and improve the model relevance. The ice volume difference between the reference issued from GPR measurements and the predictions by the proposed neural network model varies between 0.5% and 16% of the reference volume. Larger ice volume difference is mainly related to over-deepening of the bedrock resulting from past larger extent of the glacier, which information is not included in the data.

Machine learning has demonstrated potentiality for challenging physical tasks, such as inverting complex mechanisms with important data limitations. It is now competing with traditional methods that involve statistical and physical modeling. These methods face significant challenges, including long computation time, extensive prior knowledge requirements, and sensitivity to scarce and noisy data which limit their ability to generalize. Regarding these difficulties, this article aims to explore the potential deployment of a deep learning-based method to solve an inverse problem in volcanology, that is, to estimate the volume change and depth of a Mogi-type source model from surface displacement measurements. Simulated displacement samples are used to get rid of insufficient amounts of real data and a lack of ground truth. Particular efforts are devoted to proper data preparation, including proposing a semi-automatic technique for training, validation, and testing data sampling and investigating the impact of data distribution, data diversity, and noise. Real data over the Suswa volcano are also used to further assess the performance of the proposed deep learning method. Results with both synthetic and real data provide evidence to consider deep learning-based methods for geophysical inverse problems.

Glaciers play a critical role in the Earth’s climate system, and accurate estimates of their behaviours are essential for understanding the impacts of climate change and informing policy decisions. One of the most important parameters for such a task is ice distribution, which is difficult to measure and predict using traditional physics-based models. In this study, we propose a deep learning approach to predict glacier thickness by learning directly from ice velocity and topography. Our approach overcomes the limitations of traditional physics-based models, such as computational cost and the need for expert knowledge to calibrate the models. In addition, deep learning models are flexible enough to explore the relevance of multimodality and multitasking to address the physical problem. Our results demonstrate the feasibility of quickly training a neural network model with sufficient training data and producing stable, high-quality ice thickness estimates. We highlight the importance of some specific input features suggested by geophysicists that have a positive impact on model stability.

Detecting and characterizing slow slip events (SSEs) in Global Navigation Satellite System (GNSS) time series is challenging and multi-station deep-learning approaches are still little explored. The main difficulty is the high level of noise in GNSS time series. The noise affecting GNSS measurements is spatially and temporally correlated, which requires setting up multi-station methods to better characterize the spatial extent of slow slip events. Here, we develop and compare different deep-learning approaches to detect and characterize SSEs in GNSS data, showing that methods embedding the spatial information outperform time-series-based approaches, with spatiotemporal models being the most promising and flexible on real GNSS data.

Image correlation is a powerful method for remotely constraining ground displacements associated with natural disasters. By employing sub-pixel correlation algorithms, one can obtain a displacement field by correlating satellite images acquired before and after a displacement event. However, this computation may be biased when dealing with sharp discontinuities, typical of earthquake surface ruptures, which are of current interest in the context of quantifying the partitioning of slip between the primary fault core and neighboring damage zone. In this paper, we present an innovative deep learning method to perform sub-pixel correlation of optical satellite images for the retrieval of ground displacement, designed to mitigate bias around fault ruptures. From the generation of a realistic simulated database of images before and after synthetic ground displacement built specifically to deal with fault discontinuities in satellite images (e.g. Landsat-8 in this case), we developed a Convolutional Neural Network (CNN) able to retrieve sub-pixel displacements. Comparison with a state-of-the-art phase correlation method shows our pipeline is able to mitigate the sub-pixel bias in the near-field of earthquake ruptures.

Precise estimation of ground displacement at regional scales from optical satellite imagery is fundamental for the study of natural disasters, such as earthquakes, volcanoes, landslides, etc. Current methods make use of correlation techniques between two acquisitions in order to retrieve a fractional pixel shift. However, differences in local lighting conditions between two acquisitions can lead to differences in image reflectance, which in turn can bias the displacement estimate, especially in the sub-pixel domain. Data-driven methods may provide a way to overcome these errors. From the generation of a realistic simulated database based on Landsat-8 satellite image pairs with added simulated sub-pixel shifts, we developed a Convolutional Neural Network (CNN) able to retrieve sub-pixel displacements.

In this study, we develop a novel method to automatically detect areas of snow avalanche debris using a color space segmentation technique applied to synthetic aperture radar (SAR) image time series through January 2018 in the Swiss Alps. Debris avalanche zones are detected assuming that these areas are characterized by a significant and localized increase in SAR signal relative to the surrounding environment. We undertake a sensitivity study by calculating debris products by varying the D-M reference images (a stable reference image taken several weeks before the event). We examine the results according to the direction of the orbit, the characteristics of the terrain (slope, altitude, orientation), and also by evaluating the relevance of the detection with the help of an independent SPOT database by Hafner and Buhler[1]including 18 737 avalanche events. Small avalanches are not detected by SAR images, and depending on the orientation of the terrain some avalanches are not detected by either the ascending or the descending orbit. The detection results vary with the reference image; best detection results are obtained with some selected individual dates with almost 70% of verified avalanche events using the ascending orbit.

This article discusses the issue of snow avalanches monitoring at regional scale in the French Alps using SAR observations from Sentinel-1. A color space segmentation technique applied to SAR image time series has recently been used to detect areas of avalanche debris and successfully evaluated in the Swiss Alps. The objective of this paper is to generalize the detection of avalanche debris to all the French massifs and to develop indicators derived from our satellite product Sentinel-1 that can provide information about the avalanche activity at the spatial scale of the massifs. Evaluations are also planned using model outputs, in-situ measurements and very high resolution optical satellite observations.

The use of synthetic sequences is one of the most promising tools for advanced in silico evaluation of the quantification of cardiac deformation and strain through 3-D ultrasound (US) and magnetic resonance (MR) imaging. In this paper, we propose the first simulation framework which allows the generation of realistic 3-D synthetic cardiac US and MR (both cine and tagging) image sequences from the same virtual patient. A state-of-the-art electromechanical (E/M) model was exploited for simulating groundtruth cardiac motion fields ranging from healthy to various pathological cases, including both ventricular dyssynchrony and myocardial ischemia. The E/M groundtruth along with template MR/US images and physical simulators were combined in a unified framework for generating synthetic data. We efficiently merged several warping strategies to keep the full control of myocardial deformations while preserving realistic image texture. In total, we generated 18 virtual patients, each with synthetic 3-D US, cine MR, and tagged MR sequences. The simulated images were evaluated both qualitatively by showing realistic textures and quantitatively by observing myocardial intensity distributions similar to real data. In particular, the US simulation showed a smoother myocardium/background interface than the state-of-the-art. We also assessed the mechanical properties. The pathological subjects were discriminated from the healthy ones by both global indexes (ejection fraction and the global circumferential strain) and regional strain curves. The synthetic database is comprehensive in terms of both pathology and modality, and has a level of realism sufficient for validation purposes. All the 90 sequences are made publicly available to the research community via an open-access database.

Quantification of cardiac deformation and strain with 3D ultrasound takes considerable research efforts. Nevertheless, a widespread use of these techniques in clinical practice is still held back due to the lack of a solid verification process to quantify and compare performance. In this context, the use of fully synthetic sequences has become an established tool for initial in silico evaluation. Nevertheless, the realism of existing simulation techniques is still too limited to represent reliable benchmarking data. Moreover, the fact that different centers typically make use of in-house developed simulation pipelines makes a fair comparison difficult. In this context, this paper introduces a novel pipeline for the generation of synthetic 3D cardiac ultrasound image sequences. State-of-the art solutions in the fields of electromechanical modeling and ultrasound simulation are combined within an original framework that exploits a real ultrasound recording to learn and simulate realistic speckle textures. The simulated images show typical artifacts that make motion tracking in ultrasound challenging. The ground-truth displacement field is available voxelwise and is fully controlled by the electromechanical model. By progressively modifying mechanical and ultrasound parameters, the sensitivity of 3D strain algorithms to pathology and image properties can be evaluated. The proposed pipeline is used to generate an initial library of 8 sequences including healthy and pathological cases, which is made freely accessible to the research community via our project web-page.