Anomaly detection of Synthetic Aperture Radar Interferograms with semi-supervised machine learning

Estibaliz Martinez, Yosuke Aoki

The aim of this research is to detect Earth's deformation from Interferometric Synthetic Aperture Radar (InSAR) images through a semi-supervised machine learning algorithm called Least-Squares Two-sample Test (LSTT). This algorithm computes the probability distributions of two samples to assess if they belong to the same probability distribution. At the same time, it gives the divergence of these two samples. Therefore, discrimination of noise and deformation signals requires comparing the probability distribution of the samples. Our method is tested with two different datasets obtained from the ALOS-2 satellite. One is the deformation of the 2020 Taal volcano, Philippines, eruption. The other is the deformation due to the gas extraction in the Boso Peninsula, Japan, between 2017 and 2018. The first one shows obvious deformation, while the second presents a less visible deformation. Our method, tested with the Receiver Operating Characteristic curve, mostly detected the deformation correctly as an anomaly, regardless of how visually obvious the deformation is. However, our method can sometimes lead to a biased prediction when the noise level is high.