Sentinel-1 satellites, ESA’s Synthetic Aperture Radar (SAR) mission, provide continuous data from the Earth surface in weekly to biweekly time intervals. This data availability provides an unprecedented opportunity to continuously monitor the Earth surface motion in areas prone to geohazards; such as regions of high seismic and volcanic activities, with the end goal of supporting the Early Warning Systems. However, the great challenge is to derive insights from Terabytes of satellite image sequences, in a computationally-efficient and time-critical manner. We’ve risen to this challenge by designing innovative signal processing and deep learning algorithms to efficiently mine this invaluable wealth of data. This talk gives on overview of our designed solutions, as well as a demonstration of these solutions in the Tectonic and Volcanic monitoring of South America (TecVolSA) project.
300003-World Science Day For Peace And Development.pptx
Keeping a Sentinel Eye on the Volcanoes – from Space!
1. Keeping a Sentinel Eye on theVolcanoes – From Space!
Dr. HomaAnsari
German Aerospace Center (DLR)
Earth Observation Data Science
Thanks to collaborators from
German Aerospace Center (DLR) EO Data Science
German Aerospace Center (DLR) SAR signal processing
German Research Center for Geoscience (GFZ)
Technical University of Munich (TUM)
3. Keeping the Sentinel Eye On the Ground
Sentinel-1: ESA’s synthetic aperture radar (SAR) mission
Opportunity : Global monitoring of geohazards with Interferometric SAR (InSAR)
Challenge : Diving into Petabytes of Data for insight -> Efficient Big Data Processing and Mining
5. Earth Surface Deformation from SARTime Series (TS)
𝑅1
𝑡1
𝑅1
Atmosphere at 𝑡1
Earth surface at 𝑡1
𝑡1
Acquisition at
Single SAR image
6. Earth Surface Deformation from SARTime Series (TS)
𝑅1
𝑡1
𝑡2
Δ𝑅defo
Δ𝑅 = Δ𝑟deformation + Δ𝑟atmosphere + Δ𝑟scattering + Δ𝑟system noise
∗
𝑅2 = 𝑅1 + Δ𝑅
Atmosphere at 𝑡1
Atmosphere at 𝑡2
Earth surface at 𝑡1
Earth surface at 𝑡2
𝑡1 &
Acquisition at 𝑡2
Interferogram
Pairwise SAR Time Series
7. Big Data Challenge with SARTime Series
Objectives: primarily : efficient Earth surface deformation retrieval from SARTime Series
ultimately : efficient extraction of insight from the surface deformation data products
Focus on the efficiency: designing innovative signal processing and deep learning solutions
⋯
60
SAR Acquisitions
1770
Noisy Interferograms
one-year
Data Archive
4D Mapping of the Earth
Surface
TBs of data products
9. Sequential EMI: our efficient recipe for stream processing the data
Generation of
Artificial
Interferograms
Compression
of Mini-stacks
Division
of Time Series to
Mini-stacks
𝒕𝑵
𝒕𝟏
SAR Time Series
Global Systematic Earth Observation Missions
Open the Big Data challenge
* Ansari, De Zan, Bamler. “Sequential Estimator: towards efficient InSAR time series analysis,” IEEE TGRS. 2017.
** Ansari, De Zan, Bamler. “Efficient Phase Estimation for Interferogram Stacks,” IEEE TGRS. 2018.
*** Ansari. “Efficient High-Precision Time Series Analysis for SAR Interferometry,” Doctoral Dissertation, TUM+DLR, 2020
10. Deformation
Velocity
[mm/yr]
Sentinel-1 A/B
Sicily - Italy
10.2014 - 09.2018
905 Interferograms 1785 Interferograms 16836 Interferograms
Impact of Big Data Processing Recipes on Accuracy
Minimal Exploitation of TS Increased Exploitation of TS Our Efficient Recipe
* Ansari, De Zan, Bamler. “Study of Systematic Bias in Measuring Surface Deformation With SAR Interferometry,” IEEE TGRS. 2021.
11. Efficient SARTime Series Processing in Action
4D mapping of the Earth surface
SAR Image Time Series
Processing
Sentinel-1
data acquisition
5 Year SAR Data
Sentinel-1 A/B
2014 – 2019
Central Volcanic Zone of the Andes
TecVolSA: monitoringTectonics andVolcanoes of South America
13. TecVolSA: Earth Surface Deformation Products
Retrieved Data Products
❑ Surface Deformation Velocity Maps
❑ Point-wise Deformation Time Series
temporal evolution of the Earth surface
H. Volcanoes
P. Volcanoes
5 Year SAR Data
Sentinel-1 A/B
2014 – 2019
Central Volcanic Zone of the Andes
14. Diversity of Geophysical Signals
Chile - Atacama Salt Pane
Lithium mining activities
Andes - Lastarria Volcano
Volcanic deformation
Hard carbonate
crust uplifting
Soft sulphate crust
subsiding
Argentina - Tebenquicho
Solifluction signal
Chile – Illapel
2015 8.3 Mw Earthquake and aftershocks
Displacement
Velocity
[mm/yr]
+5
-5
0
Coseismic interferogram Surface Displacement Time Series
In-situ Measurement Campaign by GFZ
Surface Displacement Map
+20
-20
EQ event
16 Sept. 2015
18. The Rationale in Volcanic Deformation Detection
Binary classification
o Volcanic
o Non-volcanic
Convolutional Neural Nets
Inference/Detection: Real Data
o SAR-derived Deformation Velocity Maps
Training the networks: Simulated Data
o Volcanic deformation simulation: Geophysical models
o Atmospheric simulations: Retrieved Statistics of atmospheric noise
19. Sentinel-1 A/B
2014 – 2019
CentralVolcanic Zone of theAndes
Wide AreaVolcanic
Activity Detection
Exemplary patch-wise inference of
CNNs over the volcanic belt
Volcanic
Activity
Teo.Beker@dlr.de
22. The problem at hand
Deformation Time Series Products:
▪ Point-wise time series
▪ millions to billions of time series
The challenge:
▪ Efficient exploration of the big data
▪ Low SNR due to the atmospheric errors
▪ Arbitrary complex non-linear deformation signals
Our approach:
▪ Unsupervised machine learning (ML)
Point-Wise Deformation Time Series
4D Mapping – Multi-temporal SAR
• H. Ansari • Oct. 2021
DLR.de • Chart 22
23. The Rationale in Time Series Mining
Input
Deformation Time Series
1. Various Preprocessing
of the time series
2. Learning
Latent Features
3. Clustering
based on the Latent Features
Output
Clusters of Similar TS
* Ansari, Rußwurm, Ali, Montazeri, Parizzi, Zhu. “InSAR Displacement Time Series Mining: A Machine Learning Approach,” IGARSS 2021.
• H. Ansari • Oct. 2021
DLR.de • Chart 23
24. Various Data Compressions in the Pipeline
1. Various Preprocessing
of the time series
3. Clustering
based on the Latent Features
▪ Compressing the time series (𝑇) to latent features (𝑚)
▪ Reducing millions of points (𝑛) to hundreds of clusters (𝑘)
Output
Clusters of Similar TS
𝑛 × 𝑡
𝑛: point-wise measurements
𝑡: size of time series
𝑡 → 𝑚
𝑛 × 𝑚
𝑚: Latent Features
𝑛 → 𝑘
𝑘 × 𝑡
𝑘: Clusters
2. Learning
Latent Features
Input
Deformation Time Series
• H. Ansari • Oct. 2021
DLR.de • Chart 24
25. Deep Latent Feature Learning
Sequence Model:
▪ Transformer Encoder
𝑚
Latent
Features
Encoder Decoder
Noise-aware Transformer Autoencoder
Latent Space
Input
𝑋 ∈ 𝑅1×𝑡
Noisy TS
Output
𝑋 ∈ 𝑅1×𝑡
Deformation TS
Output
𝐿 ∈ 𝑅1×𝑚
Latent Features
Linear
Layer
Feature
Repeater
Sequence Model:
▪ Transformer Encoder
Noise Power: Γ
Encoder: Z = 𝑓Θ(𝑋)
Decoder:
𝑋 = 𝑓Θ′
′
𝑍
Noise-Aware Loss function:
ℒ =
1
2𝑡
𝑋 −
𝑋 Γ
2
+
1
2𝑡
𝑖=1
𝑡
log(Γ𝑖𝑖)
Optimization:
𝑓, 𝑓′ = argminΘ,Θ′,Γ ℒ
Outputs:
o Latent deformation learning 𝐿
o Time Series Denoising
𝑋
o Modeling noise power per epoch Γ𝑖𝑖
• H. Ansari • Oct. 2021
DLR.de • Chart 25
27. Latent Deformation Features
versus
Linear Deformation
from signal processing
using spatial information
• H. Ansari • Oct. 2021
DLR.de • Chart 27
Learned Latent Features (𝑚 = 5)
from the Noise-Aware Autoencoders
spatial information not revealed to the Autoencoders
28. Time Series Denoising and Clustering
InSAR velocity map
Clustered points shown in black dots
Clustered InSAR Time Series
(Based on the latent deformation sources)
Denoised InSAR Time Series
(Reconstructed by Noise-Aware Autoencoder)
• H. Ansari • Oct. 2021
DLR.de • Chart 28
29. In short
o Innovative algorithms are a necessity for exploiting the mission capabilities to the fullest
o Focus must be on the efficiency of algorithms but not at the cost of accuracy
o We resorted to a combination of resources to realize this objective:
o Data and dimensionality reduction
o Statistical signal processing
o Supervised learning with convolutional neural networks
o Unsupervised learning with combination of sequence models
• H. Ansari • Oct. 2021
DLR.de • Chart 29