Amazonia Biodiversity Estimation
using
Remote Sensing
and
Indigenous Taxonomy
REMOTELY SENSED DATA
We have acquired from Tropical Forests Information Center (TRFIC) several LANDSAT 5 images covering the area in object at diverse époques including high and low water seasons. The data was already geo-referenced and radiometrically calibrated
The data (bands 1-2-3-4-5-7) was transformed into GRASS raster data and opportunely rectified, registered. and patched.
Using the freely available program 6s from NASA/GFSC [27] atmospheric correction was performed to remove Rayleigh radiance and aerosol radiance using the ' dark object technique" .
The data was successively terrain corrected using the available DEM data and the GRASS program g.mapcalc.
The resulting data was used to produce Vegetation Indices and Landcover Clusters
|
Normalized Difference Vegetation Index
|
NDVI = (TM4 -TM3 / (TM4+TM3)
|
|
Transformed Normalized Difference Vegetation Index
|
TNDVI = (NDVI+0.5)/2
|
Based on the tasseled cup transformation and using a set of empirically derived coefficients [14] the following indices were obtained
|
Soil Brightness
|
= | 0.3037*TM1 + 0.2793*TM2 + 0.4743*TM3 + 0.5585*TM4 + 0.5082*TM5 + 0.1863*TM7
|
|
Vegetation Greenness
|
= | - 0.2848*TM1 - 0.2435*TM2 - 0.5436*TM3 + 0.7243*TM4 + 0.0840*TM5 - 0.1800*TM7
|
|
Vegetation Wetness
|
= | 0.1509*TM1 + 0.1973*TM2 + 0.3279*TM3 + 0.3406*TM4 - 0.7112*TM5 - 0.4572*TM7
|
Other indices obtained were:
|
Soil wetness
|
TM5-TM2
|
|
Water reflectance
|
TM2-TM4
|
|
Spectral Mixture Analysis (SMA) was developed in 1986 By Adams, Smith and Johnson fo r the Viking Mission [28].
Markus Neteler has given a very clear presentation of SMA [29].
SMA assumes that the reflectance of each pixel is a linear combination of contributing sub-pixels components.
Examples of these ground components are green vegetation, dead vegetation, soil, water, rock, etc. If the spectral signature of these components, i.e. endmembers are known the component fraction can be found by inversion and so for each pixel can be obtained the percentage of vegetation, soil, water, etc which constitute the area referred to by the pixel. The spectral signatures of the end components can be obtained from field measures or from libraries of spectra. Alternatively they can be obtained from the image itself via Principal Component Analysis and Parallel Coordinates Representation.
If the components which account for the most variance is N then the number of endmembers is N+1.
The spectral data mean corrected are projected into a N-dimensional space determined by the first N eigenvectors. to produce a N+1 vertices Polyhedron This space is called feature space or mixed space. The vertices of the polyhedron are the endmembers.
The freely available graphical analysis program XOBI provides the user with the means to explore the feature space in search of the spectra that are acceptable as the spectral signatures of ground components. XOBY offers different contemporaneous real-time views: :spectra, PCA. Tassele Cap Transformation, Spectral Ratios, etc.
The larger component areas of the landscape were singled out using the classification of the vegetation cover in the IBGE topographic maps of the area in scale 1:250000 as landcover signatures and the GRASS programs i.class. i.cluster, i.maxlik
Successively the areas thus obtained from supervised clustering were further subdivided with the help of unsupervised clustering performed using the GRASS programs i.cluster and i.maxlik. to obtain a layer of sub-patches.
Next the endemembers of these patches will be found with PCA and PCR and the patches will be analysed with SMA
Clustering applied to the resulting data will thus provide a third level patches.
This hierarchy of patches should reflect the hierarchy of the landscape ecosystems. The preliminary results appear encouraging
The patches and sub-patches thereby obtained where measured using various Fragstats metrics.
Fragstat (Spatial Pattern analysis program for quantifying Landscape Structure) was developed and released to the public domain by Kevin McGarigal and Barbara J.Marks of the Forest Science Department , Oregon State University.
The classes of metrics used include area metrics, patch density, patch size, edge metrics, shape metrics, nearest-neighbor metrics diversity metrics the latter including Shannon and Simpson indices.
We will presently be acquiring Landsat 7 data whose high resolution panchromatic band could be used for sharpening the lower resolution data.
MERIS data appears very promising and although it will have lower spatial resolution than Landsat it will have a highers pectral resolution especially in very promising zones of the spectrum. The thirteen bands provided will be radio calibrated and Rayleigh corrected but will necessitate eerosol correction. Of particular interest will be the Meris Global Vegetation Index furnished as a Level 2 product. Also very interesting will be the utilisation in tandem with ASAR and the consequent synergy which should arise. The application of SMA to the Meris data should also provide valuable new insights
SAR images are obtained from back-scattering microwave radar illumination. As a rule of thumb the rougher the reflecting surface the stronger the back scattered radiation. Also wet surfaces are brighter. because of dielectric properties.
SAR images that are acquired at different times or different vintage points can provide additional information with the help of a class of techniques referred to as SAR Interferometry Interferometric correlation , that is a measure of the variance of the interferometric phase estimate of the backscattered data obtained from two passages can provide thematic information complementary to the back-scattered intensity data.
ENVISAT ASAR has the capability to acquire data at different polarisations. In the AP mode a pair of images at different polarisations can be acquired at once. Using an algorithm based on Cloude and Pottier 's Coherent Target Decomposition [15] [16] , it is possible to extract information related to the scattering particle shape and mean orientation. We are using a implementation of Coherent Target Decomposition coded by A. A. Nielsen [17]
In the last couple of years many models have been developed in the field of SAR polarimetric interferometry which promises to separate roughness from moisture [18], to provide information on 3-D vegetation structure and biomass [19][20]. We have starting the implementation of these algorithms on GRASS and expect to be ready when ASAR data will be available.
It is increasingly recognised that the inversion of SAR data to obtain geophisical parameters should involve an initial step of segmenting the image into different terrain classes, followed by inversion using the algorithm appropriate for the particular class. We intend to do the initial segmentation using unsupervised clustering of the spectral data and bayesian methods.
JERS-1 SAR data is also interesting as it employs L-band a longer wave-length and thus is capable of penetrating the canopy more than ERS SAR and ASAR the shorter wave-length C-band.