Use of Time Series Analysis Techniques in Remote Sensing Analysis

1
Use of Time Series Analysis Techniques in Remote
Sensing Analysis

• Shruti Khanna
• CSTARS

2
Time Series Analysis

• Cyclical variation (e.g. seasonal variation)
• Measure of period, peak values, etc.
• Show seasonal changes in NDVI or temperature
• Useful in classification of land-cover, crop
  identification, NPP estimation, etc.
• Trend variation (long term change in mean)
• Estimation of trends in mean if any
• E.g. Change in NDVI in human impacted ecosystems,
  climate change detection through snow pack, etc.

3
Data Requirements

• Lots of Images !!!
• Good temporal resolution (short repeat time)
• Good temporal extent (multiyear timescale)
• All images geometrically and radiometrically
  corrected
• Should be cheap (many images required)

4
Sensors Used in Past Studies

• AVHRR 5 bands, 1 km resolution, 1 day repeat
  time, in orbit since 1987.
• MODIS 36 bands, 0.25 1 km resolution, 2 day
  repeat time, in orbit since 1998.
• SPOT Vegetation 4 bands, 1 km resolution, 1 day
  repeat time, in orbit since 1998.
• Landsat 7 bands, 30m resolution, 26 day repeat
  time, in orbit since 1972.

5
Cyclical Variation

• All ecosystems cycle due to seasonal changes in
  evapo-transpiration, day length and temperature.
• Time series analysis can characterize the period,
  amplitude and offset of the cycle.
• Grasslands, croplands, deciduous forests,
  evergreen forests, semi-arid systems, all tend to
  have different values of the above parameters and
  can be distinguished from each other (Reed et al.
  1994)
• Crops can themselves be differentiated based on
  knowledge of crop calendars and observation of
  the onset of the growing season and its length.

6
Example from current study

• Investigators Alicia, Mike, Susan, Shruti
• MODIS time series (2000 to 2004)
• 8-day composite data, 46 files
• 500m data
• Five indices calculated
• Study area California

7
Image Analyzed California
March Color Infrared Image
8
Cyclical Variation in NDVI
9
Five Indices used

• NDVI (NIR Red)/(NIR Red)
• SWIR1_2 (SWIR1 - SWIR2) / (SWIR1 SWIR2)
• NDWI_swir1 (NIR SWIR1) / (NIR SWIR1)
• NDWI_swir2 (NIR SWIR2) / (NIR SWIR2)
• GreenNDVI (Red - Green)/(Red Green)

10
Time Series Plots of Five Spectral Indices in
Rangelands
11
Some Results of Fourier Analysis

• NDWI has shorter seasonal cycle than NDVI
• NDWI maxima occurs before NDVI maxima
• NDVI responds to solar radiation and
  precipitation simultaneously

12
Trend Analysis

• Elvidge et al. 1998
• Used 64 MSS images in time series analysis of
  wetlands.
• They fit a sine function to the data to determine
  annual cycle.
• They subtracted this from the data to determine
  residuals and identify long term trends.
• Wide fluctuations in trend with climate indicated
  less stable wetlands.

13
Myneni et al. 1997
14
Myneni et al. 1997

• Used time series analysis to detect correlation
  between CO2 elevation in atmosphere and NDVI
  trends across the Globe.
• They found maximum increase in NDVI values north
  of 50 degree latitude, mostly inland from the
  ocean.
• They surmised that this was due to early snowmelt
  leading to early start of the growing season and
  increased temperatures in winter and spring.

15
Bjorgo et al. 1997
16
Bjorgo et al. 1997

• Used time series analysis of RADAR imagery (1978
  1995) to detect trends in Arctic Ice Pack.
• They merged the datasets of SMMR and SSMI
  time-series using overlap in data collection
  creating one long 16.8 year dataset which would
  otherwise not be possible.
• They found that the ice pack was shrinking as had
  been expected due to global warming.

17
Limitations of Time Series Analysis

• Myneni et al. study was criticized for not
  accounting for artifacts that result from
  comparison of multi-temporal images.
• If the images are not georegistered to sub-pixel
  accuracy, analysis can yield spurious results (or
  artifacts).
• Radiometric calibration is equally important to
  tell real differences between pixels from
  differences resulting due to different
  radiometric scales of the multiple images.

18
Limitations (cont.)

• Ecological time scale data do not exist in remote
  sensing yet. Maximum temporal extent is 30 years
  (Landsat) and satellite sensors with good
  temporal repeat time have been in orbit for
  hardly a decade.
• Climate change patterns and other ecological
  patterns of interest are not visible at such
  short time scales.
• Radiometric calibration, data normalization and
  georegistration techniques need to catch up with
  sensor data quality. Though automated methods for
  the above exist, they are not available broadly.

19
Acknowledgements

• Susan Ustin
• Mike Whiting
• Alicia Palacios
• Javier Litago