Oleksiy Kravchenko
Senior Data Scientist, Zoral Labs, Kyiv
Satellite orbits
Geostationary satellite images - MSG
Polar orbiting images - MODIS
Multispectral data
 Data
– Low resolution
• MODIS (250м)
• Proba-V (100м)
• Sentinel-3 (300 m)
– Medium resolution
• Landsat-7 (30м)
• Landsat-8 (30м)
• Sentinel-1 (20м,
2013)
• Sentinel-2 (10м,
2015)
Довжина хвилі, нм
Відбиваючаздатність
Поглинання
хлорофілу Поглинання води
Поглинання
сухої біомаси
Визначається
внутрішньою
структурою
покриву
Видимий
(VIS)
Ближній
інфрачервоний
(NIR)
Середній інфрачервоний
(SWIR)
NDVI = (NIR – RED)/
(NIR + RED)
Applications
 Agriculture
– Agricultural statistics
• Crop area estimation
– Yield forecasting
– Crop state & dynamics monitoring
– Drought monitoring
– Irrigated land monitoring
– Crop rotation control, subsides control
– Precision agriculture
• Variable prescription maps
 Forestry
– Clear-cuts mapping, species mapping
– Biomass estimation
 Land cover mapping
7
Satellite features for crop classification
L5 2011-11-08
EO1 2012-05-23
L7 2012-05-10
Sich2 2012-04-26
Temporal dynamics
Crop rotation violations (3 years period)
Rapeseed area
2011,2013: 490 ha
2011,2012: 430 ha
2012,2013: 435 ha
2013: 1540 ha
2012 2220 ha
2011 1610 ha
Lviv site, SAR data example, 2013
Landsat8
2013-06-15 Radarsat-2
2013-06-13
 No clouds problems
 Better crop
separation with SAR
than with optical data
Winter rapeseed
Maize
Gorodotsky county, Lviv region, Ukraine
Crop mapping, 2013
• Classification accuracy
– total 81 %
– Winter rapeseed: 96%
– +10% by using SAR
data
Crop area estimation
Project ЕС JRC “Crop area
estimation with satellite images
in Ukraine”, 2009-2011
Satellite data Ground data
Processing
· Orthorectification
· Segmentation
· Classification
Stratified Area
Frame Sampling
Along the road
survey
% pixels classified as cereals
%oatsingroundsurvey
SegmentsCrop field boundariesLC map
Area estimates
(pixel counting)
Data fusion
Adjustment of area
estimates
(Regression estimator)
Final results
· Area estimates
· Accuracy
assessment
Data:
MODIS
AWiFS
Landsat-5/TM
LISS-III
RapidEye
Satellite data effect:
costs decrease in 1.5 times
Area estimation
Odesa region
   regsample yVyV

rel_eff
  x yybxreg x 
 Vy
Nn
Nnn
G
n
reg
x
y()










 1
1
3
2
1
2
2
2 2

G
k
x
x
x
 3
3

 ( )Vy
nys rreg xy 
1
12 2
   regsample yVyV ˆˆrel_eff 
r = 0.986
rel_eff = 33.4
r = 0.997
rel_eff = 165.8
Data Area
th. ha
Error
(2σ)
ths. Ha
%
Sample 108.32 51.2 57.4%
MODIS 95.18 8.88 9.32%
Landsat 96.18 3.98 4.14%
Ministry
of Agric. 101,0 - -
Yield forecast
Winter wheat yield forecast, 2013
Forecast issue date – May 1st , 2013
Precision agriculture
Irrigation monitoring
August, 2012
Vegetation State Estimation (Forward &
Inverse Problems)
Leaf model
(PROSPECT)
Canopy RTM
(4SAIL)
Atmosphere RTM
(6S)
ll TR ,
3. Vegetation state
estimation using
radiometrically corrected
satellite data
Inverse problems
2. Vegetation state
estimation using satellite
data with atmospheric
correction
1. Vegetation state
estimation using in-situ
spectra
 ,, vs
soilR
lSLADLAI ,,
mwab CCCN, ,,
canopyR
TOAR
'
s
s
s
s
v
v s
dE
k E
dz
dE
a E E s E
dz
dE
a E E s E
dz
dE
K E E E w E
dz


 

 

 
 
 
     
      
        
1
( ) ( ) ( )i i ek K C k
N
   
90 1,90
,
90 1,901
N
N
N
R
R
R

 
 




 

Satellite product validation (crop state)
• Ukrainian GEO JECAM polygon
• Hemispherical photography (circular fisheye lenses)
LAI=0.22 fCover=16%
ALA=16º
LAI=4.0 fCover=79%
ALA=65º
Illegal cropping
znaydeno.com.ua
Illegal crops on 5.9 Ha, 7k hryvnas fines payed
Satellite image preprocessing.
Sentinel-2 example
Source: SPOT SATELLITE
GEOMETRY HANDBOOK, 2002
Direct model
Model parameters
• Time, location, velocity (satellite center mass), 1Hz
• Exterior attitude, 1Hz
– measured by star trackers
• Interior camera attitude within satellite coordinate system
– rotation matrix to align camera and star tracker
• Angular position of each pixel in camera coordinates
– Angular position of pixel assembly in focal plane
– Lenses distortion
Inverse model (RPC)
Direct model georeferensing of Sich2
satellite
• Geolocation error
– RapidEye (20-50 м)
– Landsat5 (30 м)
– Spot6,7 (10 м)
– Sich2 (200-700 м)
• Kyiv (240 м)
• Shatsk (550 м)
Shatsky National Park
Sich2, 2011-11-02
550m
USGS approach
for Landsat5
Drones georeferencing by telemetry
data only
Processing pipeline
Image based registration methods
Linear correlation surface Phase correlation
   
   
 
*
1 2
*
1 2
, ,
, ,
d dx y x y x x y y
x y x y
IM IM
e
IM IM
    
   


Phase correlation explained
Accuracy estimation
Linear vs. phase correlation
10-100 more control points
Odesa region
20 м ~ 6”
Sich2 image georeferencing using
Landsat data as base map
400 m
0 m
Crimea, 2012-05-19
Orthorectification
1A
1А-Orho
Zaporizhzhya, 2012-02-11
Georef. example (1)
Khersonsky region, 2012 -05-12
1A
1A-Ortho
Georef. example (2)
Pixel assembly calibration
Corrections.
Metadata
2011-10-03
Trend analysis Corrections
provided in metadata
2012-09-
05
After
launch
After
launch
1 year
on orbit
Pixel assembly coregistration (1)
1 year
on orbit
Band coregistration
PAN NIR
Along-track
shifts
DEM
SRTM
Jitter: Sich2 example
Roll oscillation
Amplitude ~ 4m
DI
Pitch oscillation
Amplitude ~ 2m
DI
DI, pixels
Jitter (2)
2011-10-05 2011-10-28
Reducing jitter
 SSTL, 2008
 Attitude Determination through
Registration of Earth Observational
Imagery
 Momentum Wheel Activation (Y-Axis) -
DU000373
Pixel assembly coregistration (1)
Sentinel-2
Pleiades
Pixel assembly coregistration (2)
Satellite image vs. aerial
registration
airplane
tracks
Distortions due to tall buildings
Questions?

DataScience Lab2017_Коррекция геометрических искажений оптических спутниковых снимков Алексей Кравченко

  • 1.
    Oleksiy Kravchenko Senior DataScientist, Zoral Labs, Kyiv
  • 2.
  • 3.
  • 4.
  • 5.
    Multispectral data  Data –Low resolution • MODIS (250м) • Proba-V (100м) • Sentinel-3 (300 m) – Medium resolution • Landsat-7 (30м) • Landsat-8 (30м) • Sentinel-1 (20м, 2013) • Sentinel-2 (10м, 2015) Довжина хвилі, нм Відбиваючаздатність Поглинання хлорофілу Поглинання води Поглинання сухої біомаси Визначається внутрішньою структурою покриву Видимий (VIS) Ближній інфрачервоний (NIR) Середній інфрачервоний (SWIR) NDVI = (NIR – RED)/ (NIR + RED)
  • 6.
    Applications  Agriculture – Agriculturalstatistics • Crop area estimation – Yield forecasting – Crop state & dynamics monitoring – Drought monitoring – Irrigated land monitoring – Crop rotation control, subsides control – Precision agriculture • Variable prescription maps  Forestry – Clear-cuts mapping, species mapping – Biomass estimation  Land cover mapping
  • 7.
    7 Satellite features forcrop classification L5 2011-11-08 EO1 2012-05-23 L7 2012-05-10 Sich2 2012-04-26
  • 8.
  • 9.
    Crop rotation violations(3 years period) Rapeseed area 2011,2013: 490 ha 2011,2012: 430 ha 2012,2013: 435 ha 2013: 1540 ha 2012 2220 ha 2011 1610 ha
  • 10.
    Lviv site, SARdata example, 2013 Landsat8 2013-06-15 Radarsat-2 2013-06-13  No clouds problems  Better crop separation with SAR than with optical data Winter rapeseed Maize
  • 11.
    Gorodotsky county, Lvivregion, Ukraine Crop mapping, 2013 • Classification accuracy – total 81 % – Winter rapeseed: 96% – +10% by using SAR data
  • 12.
    Crop area estimation ProjectЕС JRC “Crop area estimation with satellite images in Ukraine”, 2009-2011 Satellite data Ground data Processing · Orthorectification · Segmentation · Classification Stratified Area Frame Sampling Along the road survey % pixels classified as cereals %oatsingroundsurvey SegmentsCrop field boundariesLC map Area estimates (pixel counting) Data fusion Adjustment of area estimates (Regression estimator) Final results · Area estimates · Accuracy assessment Data: MODIS AWiFS Landsat-5/TM LISS-III RapidEye Satellite data effect: costs decrease in 1.5 times
  • 13.
    Area estimation Odesa region   regsample yVyV  rel_eff   x yybxreg x   Vy Nn Nnn G n reg x y()            1 1 3 2 1 2 2 2 2  G k x x x  3 3   ( )Vy nys rreg xy  1 12 2    regsample yVyV ˆˆrel_eff  r = 0.986 rel_eff = 33.4 r = 0.997 rel_eff = 165.8 Data Area th. ha Error (2σ) ths. Ha % Sample 108.32 51.2 57.4% MODIS 95.18 8.88 9.32% Landsat 96.18 3.98 4.14% Ministry of Agric. 101,0 - -
  • 14.
    Yield forecast Winter wheatyield forecast, 2013 Forecast issue date – May 1st , 2013
  • 15.
  • 16.
  • 17.
    Vegetation State Estimation(Forward & Inverse Problems) Leaf model (PROSPECT) Canopy RTM (4SAIL) Atmosphere RTM (6S) ll TR , 3. Vegetation state estimation using radiometrically corrected satellite data Inverse problems 2. Vegetation state estimation using satellite data with atmospheric correction 1. Vegetation state estimation using in-situ spectra  ,, vs soilR lSLADLAI ,, mwab CCCN, ,, canopyR TOAR ' s s s s v v s dE k E dz dE a E E s E dz dE a E E s E dz dE K E E E w E dz                                     1 ( ) ( ) ( )i i ek K C k N     90 1,90 , 90 1,901 N N N R R R            
  • 18.
    Satellite product validation(crop state) • Ukrainian GEO JECAM polygon • Hemispherical photography (circular fisheye lenses) LAI=0.22 fCover=16% ALA=16º LAI=4.0 fCover=79% ALA=65º
  • 19.
    Illegal cropping znaydeno.com.ua Illegal cropson 5.9 Ha, 7k hryvnas fines payed
  • 20.
  • 21.
    Source: SPOT SATELLITE GEOMETRYHANDBOOK, 2002 Direct model
  • 22.
    Model parameters • Time,location, velocity (satellite center mass), 1Hz • Exterior attitude, 1Hz – measured by star trackers • Interior camera attitude within satellite coordinate system – rotation matrix to align camera and star tracker • Angular position of each pixel in camera coordinates – Angular position of pixel assembly in focal plane – Lenses distortion
  • 23.
  • 24.
    Direct model georeferensingof Sich2 satellite • Geolocation error – RapidEye (20-50 м) – Landsat5 (30 м) – Spot6,7 (10 м) – Sich2 (200-700 м) • Kyiv (240 м) • Shatsk (550 м) Shatsky National Park Sich2, 2011-11-02 550m USGS approach for Landsat5
  • 25.
    Drones georeferencing bytelemetry data only
  • 26.
  • 27.
    Image based registrationmethods Linear correlation surface Phase correlation           * 1 2 * 1 2 , , , , d dx y x y x x y y x y x y IM IM e IM IM           
  • 28.
  • 40.
  • 41.
    Linear vs. phasecorrelation 10-100 more control points
  • 42.
    Odesa region 20 м~ 6” Sich2 image georeferencing using Landsat data as base map
  • 43.
    400 m 0 m Crimea,2012-05-19 Orthorectification
  • 44.
  • 45.
    Khersonsky region, 2012-05-12 1A 1A-Ortho Georef. example (2)
  • 46.
  • 47.
    Trend analysis Corrections providedin metadata 2012-09- 05 After launch After launch 1 year on orbit Pixel assembly coregistration (1) 1 year on orbit
  • 48.
  • 49.
    Jitter: Sich2 example Rolloscillation Amplitude ~ 4m DI Pitch oscillation Amplitude ~ 2m DI DI, pixels
  • 50.
  • 51.
    Reducing jitter  SSTL,2008  Attitude Determination through Registration of Earth Observational Imagery  Momentum Wheel Activation (Y-Axis) - DU000373
  • 52.
    Pixel assembly coregistration(1) Sentinel-2 Pleiades
  • 53.
  • 54.
    Satellite image vs.aerial registration airplane tracks Distortions due to tall buildings
  • 55.