Denoising of InSAR Time Series Through Spatiotemporal Attentive Convolutional U‐Net

Abstract Over the past three decades, synthetic aperture radar (SAR) and SAR interferometry (InSAR) have revolutionized Earth observation, allowing for geophysical monitoring of Earth's surface deformation with centimeter‐to‐millimeter precision. InSAR time series are successive maps of ground motion used for natural hazards assessment with the monitoring of processes slowly deforming the Earth's surface, such as landslides, subsidence, or slow tectonic motion. In particular, the detection of transient displacements of millimeter amplitude is of utmost importance to better capture the dynamics of active tectonic faults, for a better understanding of the underlying physics governing the earthquake cycle. However, despite tremendous advances over the last two decades, noise affecting the data, primarily due to atmospheric delay, still challenges the detection of small deformation signals in InSAR time series. Consequently, an unknown amount of slow tectonic deformation signals remains concealed within the noise, raising the need for novel methodologies for signal extraction. We introduce InSARDenoiser, a spatiotemporal attentive convolutional U‐Net, designed to extract small‐scale deformation in noisy InSAR time series. The combination of a deep spatial U‐Net with a spatiotemporal Transformer tracks transient deformation in space and follows its temporal signature through time. We develop a mathematical formulation to generate realistic multi‐scale atmospheric noise and build a synthetic data set. When applied to a real InSAR time series along the North Anatolian Fault, the method effectively extracts millimeter‐scale slow deformation. The extracted deformation aligns with independent creepmeter measurements, thereby validating our method and opening the possibility of its application to diverse tectonic settings globally.