pushing modis to the edge: high-resolution applications of moderate-resolution data

Pushing MODIS to the edge:High-resolution applications of moderate-resolution data

Dániel Kristóf – Róbert Pataki – András Kolesár – Tamás Kovács

MultiTemp 2013 | Banff, AB, Canada, 25-27 June 2013

Institute of Geodesy, Cartography and Remote Sensing

Budapest, Hungary

2

Outline MODIS at a glance Beginning

Expectations and observations Particularities of MODIS dataNew preprocessing approach & first results

Current stateTime series construction and quality measuresResults achieved so far

Conclusions and outlook

3

MODIS Moderate Resolution Imaging Spectroradiometer Onboard NASA’s Terra&Aqua satellites 36 spectral bands between 0.405 and 14.385 micrometers Wide field of view, daily (or even more frequent) coverage

with nominal nadir resolutions of 250, 500 and 1000 meters Sophisticated operational data processing (MODAPS) A large number of preprocessed data & scientific products

available (for free…), timeliness Well-published algorithms, systematic revision and

reprocessing (also backwards processing; Currently: „Collection 5”)

An archive continuous in time: (almost) gapless data since 1999!

Band 2

Band 2, QC 4096(OK)

A user wants to detect the date of interventions on agricultural fields, size comparable to a 250-m MODIS pixel

4

Beginning: initial motivation…

5

Beginning: initial motivation… Aqua MODIS land

surface reflectance

Day x

Day

x+

1

Day

x+

1

Day x

Terra MODIS land surface reflectance

6

• Terra & Aqua MODIS, the same day

Terra MODIS

Aqu

a M

OD

IS

Beginning: initial motivation…

7

– Gridding artifacts: • Data stored in a predefined grid, resampled• Average overlap between observations and

their respective grid cells less than 30% [Tan et al., RSE 105(2006):98-114]

– Problems with the raster data model itself:• Fixed size and orientation of the cells although

the observation dimensions vary across the scene due to the wide field of view (+/- 55 degrees), orientation definitely not N-S

• The grid cells / pixels do not represent the area where the signal is originated from

Why…? MODIS particularities:

8Close to nadir Close to swath edge

MODIS particularities

9

Proposed solution• A possible solution would be the spatial and/or temporal

compositing and/or filtering, but: loss of resolution and information…

• Our approach: Change data representation– Calculate and store observation footprints in polygon format– Each polygon represents the respective observation footprint (“real

pixel”) sensed during image acquisition• Geolocation datasets (MOD/MYD03) contain all necessary

info to do this– Ground location, dimensions and orientation of each MODIS pixel

footprint can be determined from: Latitude, Longitude, Height, Sensor Zenith Angle, Sensor Azimuth Angle, Slant Range

– Geolocation accuracy: 50 m at 1 sigma at nadir– Swath images (MOD/MYD02) or „backsampled” Surface

Reflectance

10

Proposed solution

What does itlook like?

1 km250 m1 km

11

First results: correlation with SPOT data (NIR band)

MODIS at 1 km resolution MODIS at 250 m resolution

R2 = 0.4702 R2 = 0.6117

R2 = 0.7910 R2 = 0.7918

R2 = 0.4702 R2 = 0.6117

12

First results• Possible application: one-step radiometric

normalization of high-resolution imagery by using same-day MODIS surface reflectance.

• What next? How to handle ever-changing observation geometries as a time series?

13

TIME SERIES CONSTRUCTIONAnd then…?

14

So we selected a study area in SE Hungary…

SPOT4 image, 06/08/2003

15

So we selected a study area in SE Hungary…

SPOT4 image NIR band, 06/08/2003

16

…downloaded some MODIS geolocation files and calculated 250-m geometries…

MODIS geometry, 2003216_0945

17

MODIS geometry, 2003217_0850

…downloaded some MODIS geolocation files and calculated 250-m geometries…

18

MODIS geometry, 2003217_1030

…downloaded some MODIS geolocation files and calculated 250-m geometries…

19

MODIS geometry, 2003218_0935

…downloaded some MODIS geolocation files and calculated 250-m geometries…

20

MODIS geometry, 2003219_1020

…downloaded some MODIS geolocation files and calculated 250-m geometries…

21

MODIS geometry, 2003220_0925

…downloaded some MODIS geolocation files and calculated 250-m geometries…

22

MODIS geometry, 2003221_1005

…downloaded some MODIS geolocation files and calculated 250-m geometries…

23

MODIS geometry, 2003222_0910

…downloaded some MODIS geolocation files and calculated 250-m geometries…

24

MODIS geometry, 2003223_0955

…downloaded some MODIS geolocation files and calculated 250-m geometries…

25

MODIS geometry, 2003224_0900

…downloaded some MODIS geolocation files and calculated 250-m geometries…

26

MODIS geometry, 2003224_1035

…downloaded some MODIS geolocation files and calculated 250-m geometries…

27

…then simulated „pure” MODIS data as the zonal mean of SPOT NIR pixels…

MODIS footprints on SPOT background, 2003216_0945

28

…then simulated „pure” MODIS data as the zonal mean of SPOT NIR pixels…

MODIS footprints with SPOT-derived values, 2003216_0945

29

MODIS footprints on SPOT background, 2003217_0850

…then simulated „pure” MODIS data as the zonal mean of SPOT NIR pixels…

30

MODIS footprints with SPOT-derived values, 2003217_0850

…then simulated „pure” MODIS data as the zonal mean of SPOT NIR pixels…

31

THE GRIDDING PROCESSSetting the baseline:

32

SPOT4 image NIR band, 06/08/2003

Simulated MODIS data was then resampled according to „simple gridding” (NN)

33

SPOT4 image & predefined sinusoidal MODIS grid

Simulated MODIS data was then resampled according to „simple gridding” (NN)

34

Simulated MODIS data was then resampled according to „simple gridding” (NN)

Simulated MODIS observations, geometry: 2003216_0945

35

Simulated MODIS data was then resampled according to „simple gridding” (NN)

Simulated MODIS observations, geometry: 2003217_0850

36

Zoom-in: grid cell vs. observation geometry

37

Zoom-in: grid cell vs. observation geometry

38

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003216_0945 and the corresponding grid cell

39

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003217_0850 and the corresponding grid cell

40

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003218_0935 and the corresponding grid cell

41

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003219_1020 and the corresponding grid cell

42

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003220_0925 and the corresponding grid cell

43

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003221_1005 and the corresponding grid cell

44

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003222_0910 and the corresponding grid cell

45

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003223_0955 and the corresponding grid cell

46

Zoom-in: grid cell vs. observation geometry

MODIS observation from 2003224_0900 and the corresponding grid cell

47

Zoom-in: grid cell vs. observation geometry

All MODIS observation geometries corresponding to the selected grid cell

48

SURFACE OBJECTS AS OBSERVATION UNITS

The other approach:

49

50

Land surface objects of interest:observation units defined a priori

Crop map: 512 crop parcels delineated on HR data

51

Land surface objects of interest:observation units defined a priori

Crop map: 512 crop parcels delineated on HR data

52

Relevance („purity”) of each MODIS observation (~pixel) to each observation unit

Portion of observation belonging to the given parcel

53

Relevance („purity”) of each MODIS observation (~pixel) to each observation unit

Portion of observation belonging to the given parcel

54

Only „pure” observations (above threshold) are selected

Green: retained, Red: rejected

55

Only „pure” observations (above threshold) are selected

Green: retained, Red: rejected

56

Then, radiometric values are assigned to each observation unit (parcel) for each MODIS overpass

Calculation based on area-weigthed mean of retained pixels

57

QUALITY ASSESSMENT

58

0

10

20

30

40

50

60

70

mean_spot

mean_grid_simple

mean_modis_limit

mean_modis

0102030405060708090

100

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

105

110

115

120

125

130

135

140

145

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

46

48

50

52

54

56

58

60

62

64

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

0

20

40

60

80

100

120

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

0

20

40

60

80

100

120

mean_spot

mean_grid_simple

atlag_gridbol

mean_modis

mean_modis_limit0

20

40

60

80

100

120

140

160

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

0

20

40

60

80

100

120

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

0

20

40

60

80

100

120

140

mean_spot

mean_grid_simple

mean_modis

mean_modis_limit

0102030405060708090

100

mean_spot

mean_grid_simple

atlag_gridbol

mean_modis

mean_modis_limit

Quality assessment: time-series comparison

59

Observation time

Number of parcels

Parcels with valid obs.

Correlation gridded vs. reference

Correlation all MODIS vs. reference

Correlation filtered MODIS vs. reference

Mean diff (~RMSE) gridded

Mean diff (~RMSE) all MODIS

Mean diff (~RMSE) filtered MODIS

Max diff (~RMSE) gridded

Max diff (~RMSE) all MODIS

Max diff (~RMSE) filtered MODIS

2003216_0945 512 466 0,85 0,88 0,94 6,83 7,60 5,14 75,66 37,13 29,852003217_0850 512 333 0,73 0,74 0,89 10,60 10,71 7,72 51,57 40,87 40,182003217_1030 512 207 0,70 0,69 0,94 10,86 11,28 6,26 58,88 44,75 29,972003218_0935 512 463 0,86 0,89 0,95 6,57 7,42 4,74 63,79 41,14 36,522003219_1020 512 310 0,79 0,78 0,94 9,26 9,74 5,45 41,75 43,33 33,022003220_0925 512 460 0,85 0,87 0,93 7,39 8,01 5,59 41,89 38,60 35,212003221_1005 512 381 0,80 0,84 0,95 8,45 8,83 5,11 58,21 41,08 26,442003222_0910 512 419 0,82 0,84 0,92 8,19 8,87 6,22 58,66 43,31 31,572003223_0955 512 444 0,88 0,88 0,95 6,97 7,84 5,09 41,22 35,14 30,652003224_0900 512 363 0,74 0,77 0,91 9,83 10,05 6,86 69,21 49,45 29,012003224_1035 512 142 0,64 0,63 0,93 11,89 12,01 6,09 55,05 43,72 28,71

Quality assessment: statistics

60

Quality measure: number of retained pixels

Green: high, Yellow: medium, Orange: low, Red: zero

2003216_0945

61

Green: high, Yellow: medium, Red: low, White: N/A

Quality measure: average pixel proportion

2003216_0945

62

Quality measure: number of retained pixels

Green: high, Yellow: medium, Orange: low, Red: zero

2003224_1035

63

Quality measure: average pixel proportion

Green: high, Yellow: medium, Red: low, White: N/A

2003224_1035

64

0 500000 1000000 1500000 2000000 2500000 3000000

010

2030

40

area

diff_

spot

_mod

is

2003224_1035

0 500000 1000000 1500000 2000000 2500000 3000000

010

2030

area

diff

_spo

t_m

odis

2003226_0945

Area vs. accuracy

65

2003224_1035

40 60 80 100 120 140

6080

100

120

mean_spot

mea

n_gr

id_s

impl

e

40 60 80 100 120 140

4060

8010

012

0

mean_spot

mea

n_m

odis

_lim

it

N = 512r = 0,638RMSE = 11,89

N = 142r = 0,934RMSE = 6,09

Correlation and RMS errors

66

2003216_0945

N = 512r = 0,855RMSE = 6,82

N = 466r = 0,943RMSE = 5,13

40 60 80 100 120 140

6080

100

120

140

mean_spot

mea

n_m

odis

_lim

it

40 60 80 100 120 140

4060

8010

012

014

0

mean_spot

mea

n_gr

id_s

impl

e

Correlation and RMS errors

67

0.5 0.6 0.7 0.8 0.9

05

1015

2025

mean_pix_prop_limit

diff

_spo

t_m

odis

_lim

it

2003224_1035

5 10 15 20

05

1015

2025

pixel_count_limit

diff

_spo

t_m

odis

_lim

it

Pixel proportion, number of pixels, and RMSE

68

2003216_0945

0 10 20 30 40 50 60

05

1015

2025

30

pixel_count_limit

diff

_spo

t_m

odis

_lim

it

0.5 0.6 0.7 0.8 0.9 1.0

05

1015

2025

30

mean_pix_prop_limit

diff

_spo

t_m

odis

_lim

it

Pixel proportion, number of pixels and RMSE

69

CONCLUSIONS AND OUTLOOK

70

Conclusions• Gridding is a major source of inaccuracy and noise• New polygon representation of MODIS observations has

significantly increased correlation with same-day high-resolution data: better representation of observation geometry

• Landscape objects delineated a priori can be efficiently used to construct time series

• Data processing requires more computing power, but the user has full control over the process (thresholds, quality measures, etc.) and can obtain more accurate data with maximal spatial and temporal coverage

71

Outlook• Currently: „proof of concept” level - further work

on real case studies needed• Point Spread Function is another source of

inaccuracy – could be integrated• Other data representations (Spline surface,

SensorML, etc.) are also to be assessed• Data processing speed is already good

(Python/PostgreSQL/PostGIS), but can be further improved by e.g. parallel computing in cloud architecture: implementation of IQmulus project service

72

Thank you!

Questions?Dániel Kristóf

kristof.daniel@fomi.hu

A High-volume Fusion and Analysis Platform for Geospatial Point

Clouds, Coverages and Volumetric Data Sets

www.iqmulus.eu

IQmulus (FP7-ICT-2011-318787) is a 4-year Integrating Project (IP) in the area of Intelligent Information Management within ICT 2011.4.4 Challenge 4: Technologies for Digital Content and Languages. IQmulus started on November 1, 2012, and will finish October 31, 2016.