Pengolahan Citra RADAR

SENTINEL 1
SENTINEL-1

●
Sentinel Data Access: Sentinel 1 (and 2) imagery can be accessed via:
https://scihub.esa.int/dhus/
●
Requires user registration [once]
●
Serves all users, through different interfaces
●
But feeds from one single infrastructure
●
Combines interactive explore interface (+download)
●
and service end-point for scriptable selection
SENTINEL-1

●
Limited to 2 simultaneous downloads per account
●
(but unlimited number of accounts…)
●
Performance dependent on connection, hub load
●
Shared with > 100,000 users!
●
GUI search set-up for Sentinel-1, 2 and 3
●
Rather clunky, but OK for occasional selection
●
Beware of regular downtimes (scheduled and not)
●
Archive is still full, but will eventually be “rolling”!
SENTINEL-1

●
Non-time critical target (NTC) is < 24 hours after acquisition (green line)
●
Delays beyond NTC are due to:
– Problems at the PDGS
– Problems at the hub
●
User only experiences overall delay
●
Add download timing to get ETA
●
With current hub congestion, download timing highly variable
SENTINEL-1

●
Alternative access:
– Sentinel-hub.com (Sinergise/AWS)
– EODIAS.eu (and other DIAS instances > 20 June)!
– National collaborative ground segment (e.g. PEPS)
– Alaska SAR facility
– Google Earth Engine
– others?
●
Registration usually required, may be fee-based
●
But, all depend on ingestion in ESA data hub
●
Downloads may be lots faster, though
Sentinel-1 characteristics

●
C-band SAR, like ERS-{1|2}, ENVISAT, RADARSAT
●
185 km swath, IW mode at 10 m resolution (land)
●
12-day revisit per sensor, 6 days combined (S1A + S1B)
●
Combinations with DESCENDING (morning) and ASCENDING (late afternoon) modes
possible
●
Over Europe dual polarization (VV/VH) is default over land
Sentinel-1 characteristics

●
Ground-Range Detected (GRD) product contains backscattering intensity (VV + VH) in SAR
geometry
●
i.e. data is ordered in azimuth and range samples.
●
Each GRD image is ~ 1.6 GB per dual pol scene
●
Single Look Complex (SLC) product contains backscattering intensity + phase (VV+VH) in
SAR geometry
●
Required for interferometry (e.g. coherence)
●
Each SLC image is ~ 6 GB per dual pol scene
Optional steps in S1 processing

●
To create “special purpose” imagery:
●
Multi-look: resampling to reduce speckle (but lower resolution: 20 m or 30 m)
●
Terrain flattening: normalize effects of terrain slope, useful in (moderately) hilly terrain
●
Speckle filtering: further reduce speckle (e.g. Refined-Lee)
Essential steps in S1 processing

To create “GIS-ready” imagery (recommended order):

●
(Orbit file refinement and Thermal noise removal)
●
Calibration: convert digital numbers to backscattering coefficients (σ0)
●
Terrain correction: geocode with SRTM DEM (1 or 3 sec), required to convert to projected
imagery
●
Results in 10 m pixel spacing
Introduction to the S1 toolbox

●
Open source software tool to process Sentinel data
●
http://step.esa.int/main/download/
●
STEP covers other sensors as well
●
Implemented in Java, cross-platform
●
S1TBX provides functions needed to process SAR data
●
Bundled in SNAP, currently in version 7.0, with frequent upgrades
●
Has elaborate interface with lots of generic tools as well
Introduction to the S1 toolbox

●
Can be run interactively and as a batch tool (gpt)
●
Workflow can be defined as a graph (XML)
●
With parameter substitution on command line
●
Post-processing of SAR is a requirement (SLC or GRD)!
●
If coherence is required, use SLC!
●
Linked with scripted download to create GIS-ready SAR imagery “hands free”
Introduction to Google Earth Engine

●
Google Earth Engine created in 2010
●
Developer platform to demonstrate large scale distributed computing with co-located
massive EO data sets
●
Free for [scientific] developers
●
Hosting “free, open and redistributable” data sets
●
All MODIS, Landsat, US data sets (e.g. SRTM, NAIP)
●
Sentinel data
GEE client APIs

●
User defines workflow to select, process and render
●
Workflow can be applied to data in GEE catalogues
●
But also to uploaded [private|shared] user data
●
Rich set of image and feature processing methods
●
Rendering: producing the result of the defined workflow, e.g. on screen, export to file,
client-side evaluation
●
Workflow can be shared as scripts (not downloads!)
●
JavaScript and python APIs
GEE resources

●
You need to register for GEE, with a Google Account
●
All GEE documentation online, no books [yet]
●
Entry point: the User guide https://sites.google.com/site/earthengineapidocs/
●
Full introduction, with basic and advanced code examples, thematic tutorials, references
GEE SCRIPT

var imgVV = ee.ImageCollection('COPERNICUS/S1_GRD')

.filter(ee.Filter.listContains('transmitterReceiverPolarisation',
'VV'))
.filter(ee.Filter.eq('instrumentMode', 'IW'))
.select('VV')
.map(function(image) {
var edge = image.lt(-30.0);
var maskedImage = image.mask().and(edge.not());
return image.updateMask(maskedImage);});

var desc = imgVV.filter(ee.Filter.eq('orbitProperties_pass', 'DESCENDING'));

var asc = imgVV.filter(ee.Filter.eq('orbitProperties_pass', 'ASCENDING'));
GEE SCRIPT

var descChange = ee.Image.cat(

desc.filter(spring).mean(),
desc.filter(lateSpring).mean(),
desc.filter(summer).mean());

var ascChange = ee.Image.cat(

asc.filter(spring).mean(),
asc.filter(lateSpring).mean(),
asc.filter(summer).mean());

Map.setCenter(106.83, -6.18, 12);

Map.addLayer(ascChange, {min: -25, max: 5}, 'Multi-T Mean ASC', true);
Map.addLayer(descChange, {min: -25, max: 5}, 'Multi-T Mean DESC', true);
GEE SCRIPT

●
The Sentinel-1 mission provides data from a dual-polarization C-band Synthetic Aperture
Radar (SAR) instrument. This collection includes the S1 Ground Range Detected (GRD)
scenes, processed using the Sentinel-1 Toolbox to generate a calibrated, ortho-corrected
product. The collection is updated daily. New assets are ingested within two days after
they become available.
●
This collection contains all of the GRD scenes. Each scene has one of 3 resolutions (10, 25
or 40 meters), 4 band combinations (corresponding to scene polarization) and 3
instrument modes. Use of the collection in a mosaic context will likely require filtering
down to a homogenous set of bands and parameters.
GEE SCRIPT

●
Each scene contains either 1 or 2 out of 4 possible polarization bands, depending on the
instrument's polarization settings. The possible combinations are single band VV or HH,
and dual band VV+VH and HH+HV:
– VV: single co-polarization, vertical transmit/vertical receive
– HH: single co-polarization, horizontal transmit/horizontal receive
– VV + VH: dual-band cross-polarization, vertical transmit/horizontal receive
– HH + HV: dual-band cross-polarization, horizontal transmit/vertical receive
GEE SCRIPT

●
Each scene also includes an additional 'angle' band that contains the approximate
viewing incidence angle in degrees at every point. This band is generated by
interpolating the 'incidenceAngle' property of the 'geolocationGridPoint' gridded field
provided with each asset.
●
Each scene was pre-processed with Sentinel-1 Toolbox following steps:
– Thermal noise removal
– Radiometric calibration
– Terrain correction using SRTM 30/ASTER DEM . The final terrain-corrected values are converted
to decibels via log scaling (10*log10(x)).
●
For more information about these pre-processing steps, please refer to the
Sentinel-1 Pre-processing article.
●
This collection is computed on-the-fly. If you want to use the underlying collection
with raw power values (which is updated faster), see
COPERNICUS/S1_GRD_FLOAT.
Machine learning with S1

●
Annual declarations form a labelled feature set
●
Can be uploaded to GEE, to extract parcel averages (=reduction), e.g. from our weekly
composites
●
Structured data vectors for machine learning
●
Many ML frameworks exists, most open source.
●
We use tensorflow
Machine learning with S1

●
All NL BRP2017 (770 K features, open access Juli 2017)
●
(but also demonstrated with DK2017, BE-VL2017)
●
We select the category “Bouwland” (250K features)
●
For 8 major crops (by area %) and > 0.30 ha (175 K)
●
➢ Split in 5 sets of 20% training, 80% test samples.
●
➢ Run through tensorflow to identify “outliers”
●
➢ Less than 1 hour runtime.