vignettes/biodivMapR_02.Rmd
biodivMapR_02.RmdThe image is ready to process, parameter setting for
biodivMapR can start.
First, let’s define the (absolute or relative) path to the image file to be processed.
Here we are using the image downloaded during the first step of the tutorial. The raster file systematically needs to pair with a .hdr file which will be automatically identified.
Input_Image_File <- destfileIt is possible to provide a mask file at this first stage of the process, in order take advantage of cloud detection or any other preprocessing step, and to target specific/valid areas.
This mask file is introduced the same way as the image, by providing the absolute or relative path to the raster file.
The mask is expected to be the same spatial dimensions and extent as the raster file to process (columns and lines)
Expected values in the mask:
0 = pixel is masked
1 = pixel is selected
If possible, the mask should be written in ENVI HDR format, with 1 band and values coded in integer 8bits.
# Set to FALSE if no mask available
Input_Mask_File <- FALSEVarious files will be computed and written on the disk during the process.
These files will be written in a directory structure created within a main output directory
A sub directory named after the name of the image processed is automatically created in this main output directory, in order to facilitate identification of the results corresponding to a given image.
If the same image is processed with different parameter settings, some files may be overwritten.
Output_Dir <- '../RESULTS'If no mask file is available with the image to process, it is possible to perform basic radiometric filterings:
Elimination of non vegetated pixels: the mask can be updated based on NDVI thresholding. Pixels with NDVI lower than this threshold are masked.
Elimination of residual clouds: : the mask can be updated based on NDVI thresholding. Pixels with reflectance in blue domain higher than this threshold are masked
Elimination of shaded areas: the mask can be updated based on NIR thresholding. Pixels with reflectance in NIR domain lower than this threshold are masked.
__These radiometric filters are very basic filters and should not replace algorthms dedicated to the detection of cloud/shade/non-vegetated pixels.
NDVI_Thresh <- 0.8
Blue_Thresh <- 500
NIR_Thresh <- 1500Continuum Removal, a normalization procedure, is available when working with reflectance data. This procedure eliminates multiplicative effects, and proved to be relatively efficient in decreasing the effect of changes in illumination within individual tree crowns, when using high spatial resolution imaging spectroscopy over tropical forests.
Continuum Removal is recommended when processing high spatial resolution (spatial resolutionc \(\leq\) 10 m) multi and hyperspectral data.
Continuum_Removal <- TRUEOnce normalization is performed, the next step in
biodivMapR consists in transformation of the data using one
of the following techniques:
PCA: Principal component analysis (no rescaling of
the data, computation from covariance matrix)
SPCA: Standardized Principal component analysis (no
rescaling of the data, computation from correlation matrix)
MNF: minimum noise fraction (appropriate for high
dimensionality data such as imaging spectroscopy)
These transformations will be followed by component selection in order to perform dimensionality reduction.
TypePCA <- 'SPCA'An additional filtering step can be performed, based on the detection of outliers in components of the PCA. This strategy is computationally more demanding and significantly longer. However, this outlier detection may be useful to mask pixels including residual effect of atmosphere, or outliers which may not be relevant to include in the process of computation of spectral variation.
This PCA filtering can be performed by setting a unique parameter,
FilterPCA, to TRUE.
# PCA FILTERING: Set to TRUE if you want second filtering based on PCA outliers to be processed.
# Slower process
# Automatically set to FALSE if TypePCA = 'MNF'
FilterPCA <- FALSEbiodivMapR computes spectral diversity over a set of
pixels. The number of pixels included for each spectral diversity data
point is defined by a window size. The window size defines the spatial
resolution of all diversity maps.
The length of the side of the window is provided as input parameter:
if window_size=10, then the pixels in the spectral
diversity maps will correspond to a 10x10 pixels windows in the original
image.
Both image spatial resolution and type of ecosystem under study should be taken into account when defining the window size: in forested ecosystems, window size should be large enough to include groups of individuals.
In forested ecosystems, window size between 0.25 and 4 ha usually match with ground data and correspond to a fair compromise. However, this would need to be adapted to specific situations.
If the value of window_size is too low, this will result in low number of pixels per spatial unit, hence limited range of variation of diversity in the image.
window_size <- 10The computational resource can be adjusted:
nbCPU corresponds to the number of processing units
to be used during steps of the computation which are parallelized. Make
sure you do not ask for more than what your computer can
provide.
MaxRAM corresponds to the maximum size of file chunk
of loaded in the RAM when processing the image.
WARNING: this does not corresponds to the maximum
amount of RAM needed for data processing. Once a chunk of file is
written, its processing may require several time this amount of RAM to
perfom different processing steps. MaxRAM <- 0.5 is
reasonable when processing a 2Gb raster on a personal computer with 16Gb
RAM. However, larger values of MaxRAM may rapidly require
larger amount of RAM.
nbCPU <- 4
MaxRAM <- 0.5The main spectral diversity metrics computed by
biodivMapR are based on a clustering of normalized &
dimensionality reduced image. Following this clustering, a cluster ID is
assigned to each pixel, corresponding to the ‘spectral
species’ defined in the original publication.
The number of clusters is defined by the value of
nbclusters.
nbclusters <- 50Once parameterization has been set, the image processing can start, first with an optional radiometric filtering.