Ground-Displacement Forecasting from Satellite Image Time Series via a Koopman-Prior Autoencoder

Takayuki Shinohara

Keywords: Koopman operator, Time series InSAR, Ground Displacement, AutoEncoder

Abstract: Accurate yet lightweight forecasting of ground-displacement image sequences is key for real-time disaster mitigation and urban-planning workflows. We introduce the Koopman-Prior Autoencoder (KPA), an efficient deep-learning framework that embeds a physics-inspired prior—a Koopman operator acting along the time axis—directly into the model’s latent dynamics. A convolutional encoder first distils each SBAS-InSAR frame into a compact representation. Rather than learning arbitrary nonlinear recurrence, we constrain temporal evolution to follow a single linear operator whose spectrum is regularised to ensure stability. This Koopman prior captures the dominant, quasi-linear components of crustal deformation, enabling long-horizon predictions with orders-of-magnitude fewer parameters and FLOPs than Transformer or diffusion baselines. Trained on nationwide Japanese SBAS archives and evaluated on geographically distinct test sites (Turkey, Italy, Hawaii), KPA matches or exceeds state-of-the-art accuracy while slashing inference cost, demonstrating that a carefully chosen physical prior can unlock scalable ground-motion forecasting on modest hardware.

Submission Number: 1