Source-linked AI summary
A deep learning approach to detecting volcano deformation from satellite imagery using synthetic datasets
Nantheera Anantrasirichai, Juliet Biggs, Fabien Albino, David Bull
TL;DR
Automatic deformation detection in InSAR is hindered by severe class imbalance and atmospheric artefacts. This paper trains CNNs with synthetic interferograms combining deformation and atmospheric components, then evaluates combined synthetic-real training and targeted atmospheric correction, reducing manual inspection and false positives.
Problem
InSAR deformation detection is difficult because atmospheric artefacts can dwarf deformation signals and observed deformation samples are extremely scarce, creating severe training-data imbalance.
Method
The paper trains CNN classifiers using synthetic interferograms formed from analytic deformation models, stratified atmospheric effects, and statistically simulated turbulent atmospheric noise, with varied component combinations and weights.
Results
The combined framework reduced the number of interferograms requiring manual inspection by half and decreased false positives by >80% compared with using real interferograms alone.
Takeaways & Limitations
Synthetic examples improve generalisation, while atmospheric correction can be applied to the smaller subset of suspicious positive results for further classification.
Takeaways & Limitations
The synthetic models simplify volcano plumbing into single source geometries and assume an elastic half-space, so the framework may not detect all possible deformation patterns.
Abstract
from arXiv · showhide
Satellites enable widespread, regional or global surveillance of volcanoes and can provide the first indication of volcanic unrest or eruption. Here we consider Interferometric Synthetic Aperture Radar (InSAR), which can be employed to detect surface deformation with a strong statistical link to eruption. The ability of machine learning to automatically identify signals of interest in these large InSAR datasets has already been demonstrated, but data-driven techniques, such as convolutional neutral networks (CNN) require balanced training datasets of positive and negative signals to effectively differentiate between real deformation and noise. As only a small proportion of volcanoes are deforming and atmospheric noise is ubiquitous, the use of machine learning for detecting volcanic unrest is more challenging. In this paper, we address this problem using synthetic interferograms to train the AlexNet. The synthetic interferograms are composed of 3 parts: 1) deformation patterns based on a Monte Carlo selection of parameters for analytic forward models, 2) stratified atmospheric effects derived from weather models and 3) turbulent atmospheric effects based on statistical simulations of correlated noise. The AlexNet architecture trained with synthetic data outperforms that trained using real interferograms alone, based on classification accuracy and positive predictive value (PPV). However, the models used to generate the synthetic signals are a simplification of the natural processes, so we retrain the CNN with a combined dataset consisting of synthetic models and selected real examples, achieving a final PPV of 82%. Although applying atmospheric corrections to the entire dataset is computationally expensive, it is relatively simple to apply them to the small subset of positive results. This further improves the detection performance without a significant increase in computational burden.
1. Introduction
InSAR can support broad volcano monitoring, but atmospheric artefacts and severely imbalanced deformation examples make automatic CNN detection difficult. The paper addresses this by training CNNs with synthetic interferograms designed to distinguish deformation from atmospheric signals.
- InSAR maps surface deformation and can provide indications of volcanic unrest before eruption.
- Atmospheric artefacts can dwarf deformation signals, making threshold-based automatic detection impractical and large-scale correction time-consuming.
- CNNs previously detected rapidly deforming systems but could not reliably distinguish deformation from atmospheric artefacts in some cases.
- Only 42 of more than 30,000 global Sentinel-1 interferograms showed deformation, creating a severe imbalance for CNN training.
- Augmentation does not resolve within-class imbalance because limited deformation examples cannot represent the diversity of global volcanic deformation.
- The paper uses synthetic data comprising deformation, stratified atmosphere, and turbulent atmosphere to improve CNN discrimination.
2. Convolutional Neural Networks and Training Dataset Problems
CNNs learn image features through layered convolutional processing, but their performance depends on representative labelled data. InSAR deformation classification is especially difficult because deformation examples are rare and existing rebalancing methods may overfit limited volcanic cases.
- Convolutional Neural Networks: CNNs use convolutional layers and filters to learn hierarchical image features for classification.
- Training Dataset Problems: Imbalanced datasets skew class distributions, causing classifiers to favour majority classes and ignore minority-class features.
- Training Dataset Problems: Balancing strategies modify learning algorithms or manipulate data through downsampling and oversampling.
- Training Dataset Problems: Downsampling can discard majority-class information, while cost-sensitive learning is difficult because its cost matrix must be optimized.
- Training Dataset Problems: Only approximately 0.15% of acquired interferograms show deformation, while the global dataset contained deformation at just four volcanoes.
- Training Dataset Problems: Existing balancing approaches are unsuitable because the few known deformation patterns are not generic enough for global volcano classification.
3. Generation of Synthetic Training Data
The synthetic training data combine analytic deformation models with stratified and turbulent atmospheric simulations. Monte Carlo sampling and weather-derived atmospheric information provide varied components for constructing synthetic interferograms.
- Synthetic Components: The paper generates synthetic examples of deformation, stratified atmospheric artefacts, and turbulent atmospheric artefacts using existing models and Monte Carlo parameter selection.
- Synthetic Components: Each component dataset contains 10,000 synthetic images covering approximately 0.5° by 0.5° and 500×500 pixels.
- Synthetic Components: Figure 1 organizes deformation in columns 1–2, weak and strong stratified atmospheres in columns 3–4, and turbulent atmospheres across increasing σ2 in column 5.
- Deformation: Deformation signals use analytic elastic-half-space solutions, with Monte Carlo sampling of source parameters and satellite viewing angles.
- Atmospheric Components: Stratified atmospheric signals are modeled with GACOS using ECMWF water-vapour-delay data and iterative tropospheric decomposition.
- Atmospheric Components: Turbulent atmospheric delays are simulated as spatially correlated noise with radially symmetric, homogeneous statistical properties.
- Atmospheric Components: The turbulent-noise covariance uses maximum covariance σ2_max and decay constant κ, estimated from real Sentinel-1 interferograms.
4. Method Development
The framework trains AlexNet first on synthetic interferograms, applies it to real interferogram patches, and then retrains it with selected real and synthetic examples. Synthetic signals combine deformation and atmospheric components, while overlapping-patch probabilities are merged to classify deformation.
- Framework: The framework initially trains a CNN on synthetic images, predicts deformation probabilities for phase-jump patches, and retrains after expert review using real and synthetic examples.The prediction stage divides interferograms into overlapping patches, tests patches containing phase discontinuities, and repeats classification after retraining.
- CNN architecture: AlexNet is fine-tuned as a pretrained network, with a maximum of 50 epochs, batch size 100, and softmax outputs interpreted as deformation probabilities.AlexNet contains five convolutional and three fully connected layers; prior work found it outperformed other pretrained networks for this application.
- Synthetic data: Synthetic interferograms combine deformation, stratified atmosphere, and turbulent atmosphere components with equal or unequal weights selected from Monte Carlo-generated examples.The component combinations include D + S + T and more realistic alternatives such as D + S and D + T; examples represent difficult and easy atmospheric conditions.
- Initial evaluation: Synthetic models classify their own testing data well, with all reported accuracy and class-recall values exceeding 90%.Table 1 evaluates synthetic testing with accuracy and class recall, and evaluates the 91-class model using mean square error.
- Real-data testing: On real interferograms, combined D + S + T training performs best because its characteristics most closely resemble real signals.Combining turbulent atmosphere with deformation improves real-data performance more than combining stratified atmosphere with deformation alone.
- Retraining: Retraining with synthetic and selected real examples achieves a best positive predictive value of 82% for the αD+βS +γT model.The same retraining results report 79% PPV for the D+S +T vs S +T model and reduce false positives by more than half relative to the earlier Envisat-trained model.
5. Atmospheric correction
Atmospheric correction was applied only to the small set of positive detections because correcting all 30,000 interferograms was not feasible. It preserved true-positive detections and removed most false positives, with a remaining failure where the weather model performed poorly.
- Correcting all 30,000 interferograms was not feasible, but applying corrections to the small set of positive detections was relatively simple.The positive set included 51 results from one augmented model and two additional false positives.
- 41 true positives remained identified as positives by all three tested combination models after atmospheric correction.This indicates that correction did not deteriorate detection of the true-positive cases.
- 11 of 12 previously identified false positives were correctly rejected by all three combination models after correction.The remaining false positive was a Mount Pico interferogram with Pmax=0.652 for the three-class model.
- The Mount Pico false positive persisted because the approximately 10 km weather-model resolution performed poorly on the small island.The other two models correctly identified this interferogram after correction, with Pmax values of 0.475 and 0.076.
- Table 2 reports 12 false-positive interferograms before correction and compares their uncorrected and corrected Pmax values.Pmax is the maximum result from the “D+S +T vs S +T” and “αD+βS +γT” models.
6. Discussion
The framework addresses severe training-data imbalance by using synthetic examples, improving generalisation while retaining limitations from simplified deformation and atmospheric models. Atmospheric correction can improve detection, but computational cost and weather-model accuracy constrain its application.
- Operational scope: The framework is not yet sufficiently developed for operational global volcanic-unrest alerts.Further development is required before operational deployment.
- Synthetic training and model scope: Synthetic examples improve CNN generalisation over training with real interferograms alone, but simplified source geometries and elastic assumptions may not represent all deformation patterns.The synthetic models ignore material and rheological heterogeneities and cannot guarantee flexibility for every possible deformation pattern.
- Atmospheric correction: Correction requires computationally expensive unwrapping and depends on weather models that are less accurate near islands, so applying it only to CNN-positive results reduces overall expense without sacrificing accuracy.The correction is particularly costly in areas with low or patchy coherence.
- Other sources of error: Synthetic training omits complex atmospheric phenomena and interferometric coherence, which may cause atmospheric, vegetation, or radar-shadow signals to be misinterpreted.Examples of omitted atmospheric effects include rolls and orographic effects.
- Other sources of error: Non-atmospheric noise is learned mainly from the negative class during real-data retraining, while synthetic noise examples remain a future possibility.The authors identify synthetic examples as a potential way to expand noise representation.
7. Conclusions
The paper develops CNN frameworks that use synthetic interferograms to detect rapid deformation in large InSAR datasets. Compared with real interferograms alone, the framework reduces manual inspection and false positives, while atmospheric correction further supports classification of suspicious positives.
- 7. Conclusions: Synthetic deformation, stratified-atmosphere, and turbulent-atmosphere components address highly imbalanced training data for AlexNet classification of wrapped InSAR images.The models are initialized with synthetic datasets and retrained using selected real and synthetic examples.
- 7. Conclusions: The framework reduces the number of interferograms requiring manual inspection by half and decreases false positives by >80% versus using real interferograms alone.The framework automatically searches large volumes of wrapped InSAR images for rapid ground deformation.