Deep Learning and Wavelet Transform for InSAR Volcanic Deformation Detection

Globally, there are about 1400 active volcanoes, and each year, 20 to 50 volcanic eruptions occur, many of which lack on-site monitoring. Open-source InSAR technology, like Sentinel-1, allows tracking volcanic deformations globally, even in remote or hard-to-access locations. By utilizing persistent and distributed scatterer interferometry (PSI/DSI), InSAR data can reveal subtle, millimeter-scale deformations, enabling granular tracking of volcanic activity. Furthermore, deep learning (DL) models can automatically identify and flag these changes as an alert or for further analysis.

This experiment utilizes a classification deep learning architecture, InceptionResNet v2, to detect volcanic deformations in InSAR data. The used dataset consists of 5-year-long deformation maps covering the Central Volcanic Zone in the South American Andes and reserves the known volcanic regions for testing. The remaining data and synthetic volcanic deformations are used to train the model.

GradCAM, the explainability tool, shows that accurate identification and differentiation of deformation signals are difficult on the model due to the subtle volcanic deformations observed in InSAR data. To address this, we apply wavelet transformations and filtering techniques to enhance the data, thereby improving the performance of the deep learning model.

Applying Daubechies 2 wavelet transform emphasizes subtle large-area, mostly volcanic, signals while removing the milder high-frequency patterns. The DL models are trained, and each is tested on the data with up to four wavelet transforms. The model trained and tested on original data achieves a 64.02% AUC ROC average, while when tested on data two times transformed by wavelet transform, it improves to 84.14% AUC ROC average.

We show that Daubechies 2 wavelet transform cleans data while amplifying the volcanic deformation. A side effect is that it enlarges the small area deformations, significant in intensity. This issue can be solved by filtering the data in preprocessing. Utilizing this method, models can detect even the smallest deformations of 5 mm/year.