The document presents an approach to detecting land cover change using temporal autocorrelation analysis of hyper-temporal satellite imagery. The methodology develops an autocorrelation-based change metric to distinguish real changes from natural phenological cycles. The optimal autocorrelation lag and detection threshold are determined using simulated change datasets. When applied to MODIS imagery of settlements in Gauteng, South Africa, the approach detected over 92% of real changes with an overall accuracy of 88%.

1. An Autocorrelation Analysis Approach to
Detecting Land Cover Change using
Hyper-Temporal Time-Series Data
W. Kleynhans 1,2 B.P. Salmon 1,2 J.C. Olivier 1 K.J. Wessels 2
F. van den Bergh 2
1 Electrical and Electronic Engineering, University of Pretoria, South Africa
2 Remote Sensing Research Unit, Meraka Institute, CSIR, Pretoria, South Africa
July 25, 2011

2. Outline
Introduction
Study Area
Methodology
Results
Conclusion
Examples
www.csir.co.za

3. Introduction
Remote sensing satellite data - effective way to monitor and
evaluate land cover changes
Bi-temporal change detection approach is not always
appropriate.
The temporal frequency should be high enough to distinguish
change events from natural phenological cycles.
Having a series of equally sampled acquisitions opens the door
for the utilization basic signal processing methods.
Temporal autocorrelation was effectively used as an input
feature for a change detection alarm.
www.csir.co.za

4. Problem statement
Settlement expansion is the most pervasive form of land cover
change in South Africa.
Objective is to develop a change alarm that is able to detect
the formation of new settlement developments.
These changes typically occur in areas that are naturally
vegetated.
Change events should be distinguished from natural
phenological cycles.
www.csir.co.za

5. Outline
Introduction
Study Area
Methodology
Results
Conclusion
Examples
www.csir.co.za

6. Study Area
The Gauteng province is located in northern South Africa and
covers a total area of approximately 17000 km2
592 natural vegetation and 372 settlement no-change MODIS
pixels were identified.
181 change MODIS pixels were identified.
Each pixel has a 7 year time-series (8 day composited:
2001-2007).
www.csir.co.za

7. Change example: Quickbird image in 2002
www.csir.co.za

8. Change example: Quickbird image in 2007
www.csir.co.za

9. Hyper temporal MODIS band 4 Time Series : No change
www.csir.co.za

10. Hyper temporal MODIS band 4 Time Series : Change
www.csir.co.za

11. Outline
Introduction
Study Area
Methodology
Results
Conclusion
Examples
www.csir.co.za

12. Temporal ACF method
Assume time series in vector form:
X = Xn , n ∈ {1, 2, ..., N} (1)
The normalized ACF for time-series X:
E [(Xn − mean(X))(Xn+τ − mean(X))]
RXX (τ ) = (2)
var(X)
Stationarity assumption: mean(X) and var(X) should be constant!
Reflection time-series from nature very rarely adhere to this
assumption but some do more than others...
www.csir.co.za

13. Temporal ACF method
www.csir.co.za

14. Change metric formulation
www.csir.co.za

15. Use of simulated change to determine detection parameters
Time-series of no change areas are relatively easy to obtain
but examples of real change is much more difficult.
Solution: Obtain only examples of no-change time-series and
simulate the change time-series using time-series blending.
Using simulated change the start and rate of change can be
controlled.
Start of change date was random using a uniform distribution
Duration of change was varied between 6 and 24 months.
Using a no-change and simulated change dataset, the optimal
band lag and threshold is determined (“off-line” optimization
phase)
www.csir.co.za

16. Use of simulated change to determine detection parameters
0.9
Band 1
0.85 Band 2
Band 3
0.8 Band 4
Band 5
0.75
Band 6
Band 7
NDVI
A
0.7
O
0.65
0.6
0.55
0.5
40 80 120 160 200 240 280 320 360
Time Lag (Days)
www.csir.co.za

17. Opperational phase
A pixel is labeled as having changed by evaluating the following,
true if R b (τ ) ≥ δτ
b∗
Change =
false if R b (τ ) < δτ ,
b∗
www.csir.co.za

18. Outline
Introduction
Study Area
Methodology
Results
Conclusion
Examples
www.csir.co.za

19. Results using simulated change
Confusion Matrix, overall accuracy (OA ) and optimal threshold
(δ ∗ ) showing the best land cover change detection performance
during the off-line optimization phase using MODIS band 4 550 nm
with a lag of 96 days
Simulated No Change δ∗ OA
change (n=482)
(n=592)
Change Detected 75.17% 14.73% 0.16 80.22%
No Change Detected 24.83% 85.27%
www.csir.co.za

20. Results using real change
Confusion Matrix, overall accuracy (OA ) and threshold (δ) for the
case of real change detection using the MODIS band 4 (550 nm)
with a lag of 96 days as determined during the off-line optimization
phase
Real change No Change δ OA
(n=181) (n=482)
Change Detected 92.27% 15.35% 0.16 88.46%
No Change Detected 7.73% 84.65%
www.csir.co.za

21. Results : Timing of the change
Mean start of change OA
2001/06 70.67%
2002/06 83.57%
2003/06 85.33%
2004/06 85.43%
2005/06 84.92%
2006/06 81.74%
2007/06 76.66%
www.csir.co.za

22. Outline
Introduction
Study Area
Methodology
Results
Conclusion
Examples
www.csir.co.za

23. Concluding remarks
A change metric can be formulated by selecting the lag that
shows the highest separability between the ACF of no-change
and simulated change time-series examples.
The optimal band lag and threshold selection is determined in
an “off-line” optimization phase.
After the detection parameters are determined, the method
was run blindly over the study area.
The method shows a slight decrease in performance if the
start of change is in the first or last year.
www.csir.co.za

24. Outline
Introduction
Study Area
Methodology
Results
Conclusion
Examples
www.csir.co.za

25. Example 1 - Before
www.csir.co.za

26. Example 1 - After
www.csir.co.za

27. Example 2 - Before
www.csir.co.za

28. Example 2 - After
www.csir.co.za

29. Example 3 - Before
www.csir.co.za

30. Example 3 - After
www.csir.co.za

31. Example 4 - Before
www.csir.co.za

32. Example 4 - After
www.csir.co.za

33. Example 5 - Before
www.csir.co.za

34. Example 5 - After
www.csir.co.za

35. Example 6 - Feb 2005
www.csir.co.za

36. Example 6 - May 2006
www.csir.co.za

37. Example 6 - Sep 2006
www.csir.co.za

38. Example 6 - May 2007
www.csir.co.za

39. Example 7 - Apr 2002
www.csir.co.za

40. Example 7 - Mar 2004
www.csir.co.za

41. Example 7 - Apr 2005
www.csir.co.za

42. Example 7 - May 2005
www.csir.co.za