Change Detection for Mapping Volcanic Ash Fall Using Sentinel-1 Data

CAMILO NARANJO; PABLO EUILLADES; GUILLERMO TOYOS; LEONARDO EUILLADES; GUSTAVO VILLAROSA

Workshop; Fringe 2021; 2021

In the face of a volcanic eruption, it is important to know the affected regions to manage an emergency. Volcanic ashfall affects buildings, vegetation and harvest, water bodies, livestock and population. Aiming to identify and map the regions affected by volcanic ashfall, we adapted and applied a Temporal Decorrelation Model (TDM) [1] using Synthetic Aperture Radar (SAR) C band data acquired by the Sentinel-1 constellation.Coherence maps derived from SAR interferometry (InSAR) are useful for change detection but these maps alone can be limited for distinguishing the changes resulting from the presence of volcanic ash on the ground from changes generated by other causes such as rain, snow, wind or the seasons. To this end, the implementation of TDM aims to improve our capabilities to discriminate among both kinds of changes. The eruption of Taal volcano (Philippines) on January 12, 2020 was selected as a case study. We prepared a dataset of 93 SAR images acquired between January 02, 2017 and February 16, 2020 in interferometric wide (IW) beam mode and Level 1 Single Look Complex (SLC), polarization VV and ascending orbit. We arranged these scenes into two subsets, one that contained 89 images acquired before the eruption and another with 4 images acquired after the eruption. Subsequently, we generated 3916 pre-eruption interferometric pairs (reference pairs) and 362 post-eruption pairs. First, the total coherence is estimated for each pixel with the TDM by using as input the pre-eruption coherence maps. Temporal decorrelation changes are related to changes in the dielectric and physical structure of the scatterers and in the location where these changes occur (i.e. volume and ground). The estimation takes into account three aspects: (1) the random motion of the vegetation structure, (2) the temporally correlated dielectric changes (seasonal and normal decreasing of the coherence over time) and (3) the temporally uncorrelated dielectric changes (random). Considering these aspects, the total temporal coherence depends on three Pair-Invariant variables: (1) the Ground to Volume Ratio, (2) the Characteristic Time for Ground and (3) the Characteristic Time for Volume. The total temporal coherence for each pixel is also dependent on two Pair-Variant random variables: (1) the Random Dielectric Changes of Ground, (2) the Random Dielectric Changes of Volume. Thus, the number of model parameters becomes twice the total number of interferometric pairs (N) plus three Pair-Invariant variables (2N + 3); in this sense, coherence depending on the land cover, the properties of the vegetation and the dominant scattering. [1]Second, the extracted model parameters are used to estimate the affected area. A Cumulative Distribution Function (CDF) is constructed by using the random dielectric changes of ground and volume variables for each pixel. This CDF is the core for the estimation because it contains statistical information on how natural occurrence or randomness affect the coherence of each pixel [1]. Third, the whole procedure explained above is implemented again but with the post-eruption subset in order to generate a new post-eruption CDF. Finally, both CDFs are compared in order to assess their similarity and subsequently estimate the probability that each pixel has been affected by ashfall. Thus, a probability map is generated, where pixels with probabilities closer to one indicate the presence of volcanic ash on the ground. Our implementation of the TDM resulted in a probability map that shows higher values that are in agreement with the distribution of volcanic ash observed on the ground by the volcano observatory of the Philippine Institute of Volcanology and Seismology (https://www.phivolcs.dost.gov.ph/) and with the dispersion of volcanic ash in the atmosphere observed from Himawari-8 satellite data. However, areas where coupled effects occur, like crustal deformation and ashfall, are more complex to interpret. Subsequent work will aim at correlating the probabilities with the presence and thickness of ashfall deposits and at implementing the TDM with both C and L Band SAR data.