1. "Seasonal variability in spectral reflectance of grasslands along a dry-mesic gradient in Switzerland" Achilleas Psomas1,2, Niklaus E. Zimmermann1, Mathias Kneubühler2, Tobias Kellenberger2, Klaus Itten2 1.Swiss Federal Research Institute WSL, 2. Remote Sensing Laboratories (RSL), University of Zurich
April 29th,2005
Warsaw University

2. Overview Introduction Objectives Data Processing-Statistical analysis
Initial Results
Discussion

3. Introduction
Dry meadows and pastures in Switzerland are species-rich habitats resulting from a traditional agricultural land use.
40% of plant and over 50% of animal species present on dry meadows are classified as endangered
90% of dry grasslands have been transformed to other land cover types
TWW Project "Dry Grassland in Switzerland"(Trockenwiesen und –weiden,1995)
Creation of a federal inventory so ecologically valuable grasslands could be given an increased protection by law.

4. General Objective
To develop, apply, and test different methods based on remote sensing datasets and techniques for identification and monitoring of dry meadows and pastures in Switzerland
Main project parts:
Part A: Field Spectrometry-(Plot to Field)
Part B: Imaging Spectrometry-(Field to Region)
Part C: Multitemporal Landsat TM approach-(Region to
Landscape)

5. Field Spectrometry Objectives
Examine the potential of using the seasonal variability in spectral
reflectance for discriminating dry meadows and pastures.
Identify the best spectral wavelengths to discriminating grasslands
of different type. Which are the spectral wavelengths with statistical
significant differences?
Identify the optimal time or times during the growing season
for discriminating and classifying different types of grasslands.

6. Example of grasslands and pastures
Semi-dry [AEMB]
Dry [MB]

7. Preprocessing-Statistical analysis

8. Structure of dataset Collection-Temporal resolution
Field spectroradiometer, Analytical Spectral Devices FieldSpec Pro
4 grassland types examined along a dry-mesic gradient
12 sample fields at Aargau and Chur
12 time steps March-October
spectral signatures collected

9. Processing of data Removal of errors mentioned at the field protocol.
Identification of potentially false recordings. Changing weather-moisture conditions. Unforced errors.
Normalization of data : Continuum Removal.
Mann-Whitney U Test (Wilcox test)
Classification and Regression Tree Analysis (C&RT) on statistically significant wavelength for selection of wavelengths.
Feature space distance analysis

10. Identification of potential errors

11. Continuum Removal I
Continuum removal standardizes reflectance spectra to allow comparison of absorption features.
Spectral absorption-depth method for identifying chlorophyll, water, cellulose, lignin image spectral features
Minimization of factors like atmospheric absorption, soil exposure, other absorbers in the leaf (Kruse et al. 1985; Clark et al. 1987; Kruse et al. 1993a).
A continuum is formed by fitting straight line segments between the maxima of the spectral curve

12. Continuum Removal III

13. Continuum Removal III

14. Statistical Analysis I
Statistical significance of spectral response was tested with the Mann-Whitney U Test (Wilcox test) for a p<0.01 for each wavelength of each field per for recording day.
Analysis was done between individual fields and between each grassland type. (for every individual day)
Both the continuum removed spectra and the original recordings were tested.
Classification and Regression Tree Analysis (C&RT) on statistically significant wavelength for selection of wavelengths.
Repeated (15x) 10-fold cross validation to optimize the pruning of the tree
Feature space distance analysis

15. Classification and Regression Trees (C&RT)
Results presented on a tree are easily summarized and interpreted.
Flexible in handling different response data types and a big number of explanatory variables.
Ease and robustness of construction.
Tree methods are nonparametric and nonlinear

16. Statistical Analysis II
AEMB MB
p-value
Wavelengths
350nm x 100
351nm x 100 ..
2500nm x 100
Wavelengths
350nm x 120
351nm x 120 ..
2500nm x 120
0.002
0.038 ..
0.0004
Wilcox test
Wilcox test
For every day all possible field combination are checked for statistical significance per wavelength.
E.g.: Recording day with 6 fields (AE,AEMB1,AEMB2,MB1,MB2,MB3)
Possible combinations : 15
Significance tests: 15 combinations x 2000 Wavelengths (variables)

17. Statistical Analysis II

18. Preliminary results Details 3 Types AE: Mesic, nutrient-rich grassland
AEMB: Less Mesic, species-rich grassland
MB: Semi-dry, species-rich grassland
Aarau
9 time steps
25. Mai 10. Jun 25. Jun 21. Jul
28. Jul 15. Aug 23. Aug 02. Sep 18. Sep

19. Significant Wavelengths I

20. Significant Wavelengths II
AE AEMB MB
 
Mesic Dry

21. C&RT Analysis I C&RT for Original spectral recordings - 10th June 2004
Classification tree: Variables actually used in tree construction: [1] "b658" "b690" "b1608" "b505" "b705" "b551" "b1441" Number of terminal nodes: 8 Residual mean deviance: = / 574 Misclassification error rate: = 45 / 582
C&RT for Original spectral recordings - 10th June 2004

22. C&RT Analysis II: Misclassification error rate

23. C&RT Analysis III: Selected Wavelengths

24.  -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 
Feature space distance
Jeffries-Matusita Distance
AE AEMB MB
 
Mesic Dry

25. Discussion
Increased spectral resolution of hyperspectral recordings provide big opportunities for discriminating grassland types.
Multitemporal recordings give a better understanding of the spectral differences between grassland types during the growing season and increase the possibilities for a successful discrimination.
Continuum removed spectra gave a smaller number of significant wavelengths but overall better class separability results during the season.
C&RT enabled the dimension reduction of the hyperspectral data
Processing of the data, statistical analysis and C&RT analysis was done totally in R, easily reproducible and adjustable.

26. Thank you for your attention…

27. Feature space distance
Bhattacharyya Distance

28.
AEMB2
MB2

29. Preliminary results

30. Preliminary results

31. Preliminary results

32. Spectral Reflectance - I
The total amount of radiation that strikes an object is referred to as the incident radiation
incident radiation = reflected radiation + absorbed radiation + transmitted radiation

33. Spectral Reflectance - II
Spectral reflectance is the portion of incident radiation that is reflected by a non-transparent surface

34. Spectral Reflectance - III
Spectral Response of Vegetation

35. Field Spectrometry II

36. Scaling-I Part Sensor Spatial Resolution Spectral Resolution
Spatial Coverage
Altitude
A) Field Spectrometry
ASD Field
Spectroradiometer
0.5m
2150 bands
6-8 fields/day
1.5m
B) Imaging Spectrometry
HyMap 5m
128 bands
12km x 4km
3km
C) Multitemporal Landsat TM
Landsat TM
30m
7 bands
180km x 180km
700km

37. Continuum Removal II

38. Scaling-II

39. Preliminary results

40. Preliminary results

41. Additional

42. Additional

43. Continuum Removal I
Trees can be used for interactive exploration and for description and prediction of patterns and processes. Advantages of trees include:
(1) the flexibility to handle a broad range of response types, including numeric, categorical, ratings, and survival data; (2) invariance to monotonic transformations of the explanatory variables;
(3) ease and robustness of construction;
(4) ease of interpretation;
(5) the ability to handle missing values in both response and explanatory variables.
Thus, trees complement or represent an alternative to many traditional statistical techniques, including multiple regression, analysis of variance, logistic regression, log-linear models, linear discriminant analysis, and survival models.

44. Discussion-Further steps
Separability analysis: Euclidean ,Jeffries-Matusita, Bhattacharyya distance
Perform CART tree analysis using the statistically significant spectral bands.
Upscaling the results of the analysis to HyMap sensor .(5m spatial resolution,128bands spectral resolution).