Normalizing flow-based deep variational Bayesian network for seismic multi-hazards and impacts estimation from InSAR imagery

Wei MaarXiv 2023 · DOI · arXiv

TL;DR

This paper introduces a novel deep Bayesian network that uses normalizing flows to jointly estimate multiple cascading seismic hazards and their impacts from InSAR imagery. By modeling complex causal relationships and leveraging prior knowledge, this approach offers more accurate and interpretable disaster assessments for improved post-disaster response.

What Problem Does This Paper Address?

Following major disasters like earthquakes, rapid and accurate estimation of cascading hazards (e.g., landslides, liquefaction) and their impacts (e.g., building damage) is crucial for effective response. While InSAR imagery provides high-resolution data, existing methods often struggle with accuracy because they predict only a single hazard type. This overlooks the intricate, interconnected nature of natural disasters, where one event can trigger a chain reaction of others, leading to complex and noisy signals in the imagery. Accurately modeling these cascading processes and disentangling their individual contributions from irrelevant environmental changes remains a significant challenge.

Methods

The core innovation is a deep variational Bayesian network (BN) augmented with normalizing flows.

  • Causal Graph BN: The authors first construct a generalized causal graph representing the dependencies between various hazards, impacts, and environmental factors. This graph is then translated into a causal Bayesian network, where edges signify causal connections.
  • Normalizing Flows: To capture the complex, non-linear relationships between parent and child nodes in the BN, normalizing flows are employed. These are invertible, smooth mapping functions that allow for the construction of highly expressive probability distributions. This is a key departure from traditional methods that might use simpler linear combinations.
  • Variational Inference: To infer the posteriors of unobserved variables (hazards, impacts, causal coefficients) without ground truth, a stochastic variational inference algorithm is developed. This converts the inference problem into an optimization problem, aiming to approximate the true posterior with a tractable distribution.
  • Triplet Loss: To combat potential overfitting due to the expressiveness of normalizing flows and the lack of ground truth, a triplet loss is introduced. This loss function penalizes divergence between the approximated posterior and prior physics-based estimates, encouraging the model to learn more identifiable and physically meaningful representations.

Datasets

  • Real-world InSAR Imagery: Used for evaluating the model's performance.
  • 2021 Haiti Earthquake Data: Specifically, Damage Proxy Maps (DPMs) generated by the ARIA system and shaking/ground failure products from the USGS.
  • Ground Truth Inventories: Landslide and building damage inventories from StEER and GEER teams were used to synthesize ground truth data for evaluation.

The paper does not introduce a new benchmark dataset but rather uses existing data to evaluate their novel approach against baselines.

Key Insights and Ideas

  • Cascading Nature is Key: The paper strongly emphasizes that accurately modeling the interconnected, cascading nature of disasters is essential for improving hazard and impact estimation, moving beyond single-hazard approaches.
  • Normalizing Flows for Expressiveness: Using normalizing flows is a powerful technique to represent the complex, non-linear causal dependencies in disaster scenarios, offering a more nuanced understanding than simpler models.
  • Identifiability through Prior Knowledge: The triplet loss, which leverages prior physics-based models, is crucial for ensuring that the highly expressive normalizing flow models are not only accurate but also interpretable and avoid overfitting.
  • Joint Estimation Advantage: The framework's ability to jointly estimate multiple unobserved hazards and impacts, along with their causal relationships, provides a more holistic and accurate picture of disaster consequences.

Key Results

  • Improved Accuracy: The proposed model significantly outperforms baseline methods, including prior physics-based models and the VBCI model, on real-world InSAR data.
  • Building Damage AUC: The building damage model achieved an AUC value of 0.9786, representing a 20.52% improvement over the prior AUC value (0.8120).
  • Qualitative Improvements: Visualizations show that as the complexity of the normalizing flow model increases (higher K), the synthesized DPMs better capture details from the real-world DPMs and ground truth. The authors' model, in particular, reproduces most details from the real-world DPMs.

Limitations and Future Work

  • Identifiability Challenges: While the triplet loss helps, the inherent complexity of deep generative models and the lack of extensive ground truth can still pose identifiability challenges.
  • Computational Cost: The use of normalizing flows and variational inference can be computationally intensive, though the authors mention developing stochastic optimization for efficiency.
  • Future Work: The authors suggest further exploration of more expressive variational inference algorithms and potentially incorporating additional data sources to further enhance the model's accuracy and robustness.

Summary updated

Abstract

Onsite disasters like earthquakes can trigger cascading hazards and impacts, such as landslides and infrastructure damage, leading to catastrophic losses; thus, rapid and accurate estimates are crucial for timely and effective post-disaster responses. Interferometric Synthetic aperture radar (InSAR) data is important in providing high-resolution onsite information for rapid hazard estimation. Most recent methods using InSAR imagery signals predict a single type of hazard and thus often suffer low accuracy due to noisy and complex signals induced by co-located hazards, impacts, and irrelevant environmental changes (e.g., vegetation changes, human activities). We introduce a novel stochastic variational inference with normalizing flows derived to jointly approximate posteriors of multiple unobserved hazards and impacts from noisy InSAR imagery.
BibTeX
@misc{ma2023normalizing,
  author = {Wei Ma},
  title = {{Normalizing flow-based deep variational Bayesian network for seismic multi-hazards and impacts estimation from InSAR imagery}},
  year = {2023},
  eprint = {2310.13805},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.13805},
}
Figure