Nice straight-forward tutorial. I'm about to graduate with a B.S. in Env. Science/Geoscience and our program is primarily all GIS & remote sensing. These quick video references are great refreshers as I finish out some senior and portfolio projects. Cheers!
Hi, I have a question, Patrick. I am working on Landsat 8 but I am confused about the relation of PCA, HVS, Pansharpening, classification. I do FLAASH on Landsat 8, Classification, HSV...., but all I have done on FLAASH image. Thus, what's the relation between them? I should not do classification on HSV images?
Ali - this is a somewhat confusing and open-ended question. in image processing when you are working with a single dataset - i.e. hyperspectral or multispectral - when you apply algorithms such as FLAASH or QUAC to this dataset that change the values within the image dataset it may change the ways that other subsequent algorithms affect the (now modified) data. I do have some limited experience with ENVI/FLAASH but do not consider myself an "expert". FLAASH and QUAC are the two atmospheric correction modules in Harris/ ENVI that are often applied to data from various types of platforms and sensors typically to improve the visual quality of images affected by atmospheric scattering and absorption of surface reflected radiation. the FLAASH algorithm modifies values in the raw image using a formula dependent on the following values: the pixel surface reflectance an average surface reflectance for the pixel and a surrounding region the spherical albedo of the atmosphere the radiance back scattered by the atmosphere two coefficients that depend on atmospheric and geometric conditions but not on the surface there are several values here that are arbitrary and in some cases categorical, and hence they will affect the final modified pixel value in ways that may be dependent on the algorithm design and/or choices made by the analyst (you). regarding the relation of this atmospheric correction module to these other categories i'm going to take a pass ;-) I would say that applying proprietary atmospheric correction modules to data that has likely already been passed through various pre-processing routines may affect subsequent processes that are applied (in your case PCA, HVS transformation, pansharpening, and classification). i would go on but the worms are escaping from the can, and these are those invasive earthworms that are so wiggly they can practically jump off your fishing hook.
thank you sir for your video, In that essence I have a question, When can I conduct PCA, after atmospheric correction or before, or just right before classification ? I am using Landsat images. cheers
You should always apply atmospheric corrections to Landsat data (and for that matter, generally, any remote sensed data) before performing PCA as the algorithms that are designed for AC routines for Landsat data normally assume they are being applied to 8- or 16-bit scaled integer numbers. in most AC routines, whether part of a system such as ENVI or applied by the user manually, it will produce DN's that are transformed into real or other 2nd generation values that may affect the efficiency of the AC routine were it to be applied after PCA.
good question - this is common in arcgis desktop when you perform an operation on a multi-raster layer that results presented such as these use the ordering of input layers that you used when loading the table. i have always somewhat taken it on faith that this is indeed what is happening. i assume you are referring to the correlation matrix? if you look at the Correlation Matrix you will notice that the r values are highest between the visible channels and also between the infrared channels. I have always observed that when you view a correlation Matrix of a landsat scene that the correlations are higher between visible channels and infrared channels, in other words visible channels are more similar to each other and less so to infrared channels. You would also normally notice that the r values are lower between visible and infrared channels, in other words visible and infrared channels are typically less similar to each other than visible to visible or infrared to infrared. This seems to be the case in this correlation Matrix. In other words it seems to confirm that the channels are listed in the order that they were entered into the table. Does that make sense?
Thanks for your response Patrick! That may be the case with satellite channel data, but in my case, I'm using gridded climate data from PRISM to do a PCA on a yearly basis...I ended up switching the order of a few of the input rasters for two different years, and it did not appear to make a difference, in the resulting matrices and the actual table of accumulated eigenvalues. If you have any ideas, or are interested, I could email you my output files for a couple of years. Cheers!
interesting. switching the order should not affect the accumulated eigenvalues but you would expect the correlation matrix to shift with the changes you made. i cut my teeth on pca in the 80s working with large arrays of census data. wish i had time to look at your stuff.
Nice straight-forward tutorial. I'm about to graduate with a B.S. in Env. Science/Geoscience and our program is primarily all GIS & remote sensing. These quick video references are great refreshers as I finish out some senior and portfolio projects. Cheers!
As I do to color the satellite image according to PC?
Hi, I have a question, Patrick.
I am working on Landsat 8 but I am confused about the relation of PCA, HVS, Pansharpening, classification. I do FLAASH on Landsat 8, Classification, HSV...., but all I have done on FLAASH image. Thus, what's the relation between them?
I should not do classification on HSV images?
Great question Ali. Let me give it some thought and I will try to have a coherent answer for you.
Ali - this is a somewhat confusing and open-ended question. in image processing when you are working with a single dataset - i.e. hyperspectral or multispectral - when you apply algorithms such as FLAASH or QUAC to this dataset that change the values within the image dataset it may change the ways that other subsequent algorithms affect the (now modified) data. I do have some limited experience with ENVI/FLAASH but do not consider myself an "expert". FLAASH and QUAC are the two atmospheric correction modules in Harris/ ENVI that are often applied to data from various types of platforms and sensors typically to improve the visual quality of images affected by atmospheric scattering and absorption of surface reflected radiation. the FLAASH algorithm modifies values in the raw image using a formula dependent on the following values:
the pixel surface reflectance
an average surface reflectance for the pixel and a surrounding region
the spherical albedo of the atmosphere
the radiance back scattered by the atmosphere
two coefficients that depend on atmospheric and geometric conditions but not on the surface
there are several values here that are arbitrary and in some cases categorical, and hence they will affect the final modified pixel value in ways that may be dependent on the algorithm design and/or choices made by the analyst (you). regarding the relation of this atmospheric correction module to these other categories i'm going to take a pass ;-) I would say that applying proprietary atmospheric correction modules to data that has likely already been passed through various pre-processing routines may affect subsequent processes that are applied (in your case PCA, HVS transformation, pansharpening, and classification). i would go on but the worms are escaping from the can, and these are those invasive earthworms that are so wiggly they can practically jump off your fishing hook.
thank you sir for your video,
In that essence I have a question,
When can I conduct PCA, after atmospheric correction or before, or just right before classification ?
I am using Landsat images.
cheers
You should always apply atmospheric corrections to Landsat data (and for that matter, generally, any remote sensed data) before performing PCA as the algorithms that are designed for AC routines for Landsat data normally assume they are being applied to 8- or 16-bit scaled integer numbers. in most AC routines, whether part of a system such as ENVI or applied by the user manually, it will produce DN's that are transformed into real or other 2nd generation values that may affect the efficiency of the AC routine were it to be applied after PCA.
so many thanks sir, it was really informative
Please, which of the bands am I combining in landsat 8?
Good question. I think it should be ok to use all but probably not channel 9 - really looking at the atmosphere.
Is there any way to be sure that Layer 1 in your PCA text file corresponds to the first raster input?
good question - this is common in arcgis desktop when you perform an operation on a multi-raster layer that results presented such as these use the ordering of input layers that you used when loading the table. i have always somewhat taken it on faith that this is indeed what is happening.
i assume you are referring to the correlation matrix? if you look at the Correlation Matrix you will notice that the r values are highest between the visible channels and also between the infrared channels. I have always observed that when you view a correlation Matrix of a landsat scene that the correlations are higher between visible channels and infrared channels, in other words visible channels are more similar to each other and less so to infrared channels.
You would also normally notice that the r values are lower between visible and infrared channels, in other words visible and infrared channels are typically less similar to each other than visible to visible or infrared to infrared.
This seems to be the case in this correlation Matrix. In other words it seems to confirm that the channels are listed in the order that they were entered into the table.
Does that make sense?
Thanks for your response Patrick! That may be the case with satellite channel data, but in my case, I'm using gridded climate data from PRISM to do a PCA on a yearly basis...I ended up switching the order of a few of the input rasters for two different years, and it did not appear to make a difference, in the resulting matrices and the actual table of accumulated eigenvalues. If you have any ideas, or are interested, I could email you my output files for a couple of years. Cheers!
interesting. switching the order should not affect the accumulated eigenvalues but you would expect the correlation matrix to shift with the changes you made. i cut my teeth on pca in the 80s working with large arrays of census data. wish i had time to look at your stuff.