Overcoming "Physics Shock" in Earth Observation A Heteroscedastic Uncertainty Framework for PINN-based Flood Inference

Background

Rapid flood extent mapping using Synthetic Aperture Radar (SAR) is crucial for disaster response, but standard Deep Learning models lack physical grounding, producing impossible predictions. Physics-Informed Neural Networks (PINNs) are designed to embed physical laws (e.g., Shallow Water Equations) directly into the model's loss function. However, their application to real-world, noisy remote sensing data has been problematic. The core challenge is enforcing rigid spatial derivatives onto latent representations that must also fit the noisy, speckle-laden data from SAR sensors, leading to training instability.

Key Points

• The Physics Shock Problem: The central innovation is identifying and naming the failure mode of traditional PINNs on real data: Physics Shock. This occurs when attempts to enforce smooth, physically-consistent gradients (from governing equations) onto a latent space designed to absorb noisy, high-frequency SAR speckle cause catastrophic gradient divergence during training.
• Uncertainty-Aware Solution: The framework addresses instability by dynamically modulating the strength of the physics loss. It uses a probabilistic model (an Attention-Gated FNO-UNet) that predicts not just the flood extent, but also the aleatoric uncertainty (sensor noise).
  • Dynamic Warm-Start: The training protocol begins by focusing on learning the data pattern before gradually introducing the physics loss, preventing early-stage gradient conflicts.
  • Negative Log-Likelihood (NLL) Objective: By optimizing a NLL loss, the model learns heteroscedastic uncertainty (varies per pixel). The physics constraint is weighted inversely to this predicted uncertainty—enforced strictly in low-uncertainty (high-confidence) areas and relaxed in high-uncertainty (noisy) regions.
• Disentangled Uncertainty via Deep Ensembles: Using ensembles of the proposed model, the work separates two types of uncertainty:
  1. Aleatoric: Intrinsic sensor noise, captured by the NLL objective within a single model.
  2. Epistemic: Model/data uncertainty (e.g., from out-of-distribution terrain), captured by the disagreement between ensemble members.
• Performance Validation: Evaluated on the Sen1Floods11 benchmark, the model achieves a +25% relative improvement in IoU over deterministic (non-probabilistic) PINN baselines, demonstrating superior accuracy and stability.

Significance

The research provides a critical bridge between theoretical PINNs and operational disaster management. It directly tackles the core practical failure—Physics Shock—that has hindered PINN adoption for Earth Observation. The probabilistic, uncertainty-aware framework offers two major operational advances:

1. Stabilized and Accurate Mapping: It delivers more reliable flood extent maps by intelligently balancing data fidelity with physical realism.
2. Actionable Confidence Metrics: By disentangling noise from ignorance, it provides highly calibrated, physically consistent confidence bounds. This allows response agencies to understand where the model is uncertain due to noise (potentially ignorable) versus where it lacks knowledge (indicating risk), enabling more robust, real-time decision-making for mitigation efforts.