I am the person who developed the 'Improved DOS' and 'COST' atmospheric corrections models (the COST model is in the paper titled 'Image Based Atmospheric Corrections --- revisited and improved').  Keep in mind that either the DOS or COST corrections can be done with a simple stretch function which all software package contain somewhere within their tool boxes of processing procedures.  You can take the model and 'hand calculate' the input DN vs the output surface spectral reflectance (for say every 20 dn levels).  Then you can use the input-output pairs as input to the stretch function and this will convert from raw or TOA dns to reflectance values.

A very critical aspect of either correction is that a 'VALID DARK OBJECT' that can be used to compute the haze values exist within the image.  At times user blindly use the minimum dn value of the image as the dark object which can lead to over correcting for haze which means that the surface reflectance values will be to low and they will under estimate the reflectance.  For more information on the use of Improved-DOS and COST techniques also look at the papers published in 1988 and 1989 and are on the research gate pub list.

Good luck with your work.

Pat

Makmie,

First of all, I suggest you review texts concerning dark object subtraction (Mather, P.M., and Koch, M. (2011) Computer Processing of Remotely-Sensed Images: An Introduction, Fourth Edition. Chichester, UK: Wiley-Blackwell, 504 pp).

Second, I would suggest that you evaluate R for landsat analysis (http://cran.r-project.org/web/packages/landsat/landsat.pdf).  Personally, I prefer to use R for Landsat analysis because 1) it's free and 2) GIS is great unless you're not familiar with GIS.

B

Better to use e-cognition software to extract objecte based image analysis from  that software. 

Some authors say that the minimum digital number to choose has to be present in at least 1000 pixels and others prefer 4000 pixels at least (it’s just to be sure that’s a real minimum of the band because sometimes the minimum value of a band is located outside the image, in the edges)…I think that (in the case you have few images) you should check visually where these minimum are located (in general corresponding to deep clear water or to shadows)

http://www.ciesin.org/docs/005-477/005-477.html

Dear Sir,

As a GIS Developer and Educator, Geographer, and having worked a lot on Multispectral Image Analysis, I would be at your disposal whenever you like. You may visit LinkedIn "Michalis Maniadakis", in order to examine if I could be useful to you.

Sincerely,

Mihael Maniadhakis,

Geographer,

Neuroscience and A.I. Mcs.

I am the person who developed the 'Improved DOS' and 'COST' atmospheric corrections models (the COST model is in the paper titled 'Image Based Atmospheric Corrections --- revisited and improved').  Keep in mind that either the DOS or COST corrections can be done with a simple stretch function which all software package contain somewhere within their tool boxes of processing procedures.  You can take the model and 'hand calculate' the input DN vs the output surface spectral reflectance (for say every 20 dn levels).  Then you can use the input-output pairs as input to the stretch function and this will convert from raw or TOA dns to reflectance values.

A very critical aspect of either correction is that a 'VALID DARK OBJECT' that can be used to compute the haze values exist within the image.  At times user blindly use the minimum dn value of the image as the dark object which can lead to over correcting for haze which means that the surface reflectance values will be to low and they will under estimate the reflectance.  For more information on the use of Improved-DOS and COST techniques also look at the papers published in 1988 and 1989 and are on the research gate pub list.

Good luck with your work.

Pat

Thank you very much to all who have given advices. I am now dealing with your advise, case by case. I ll be back once i am stuck! 

Thank you again....Brian F Thomas ,Karuturi Venkata Suryabhagavan , Nguyen Tien Hoang ,Hassane Moutahir , Mihael G. Maniadhakis,  ·Pat S. Chavez, Jr. 

How did you go with this, Makemie? I am trying to do the same thing. My minimum is clearly visible in ArcMap (in the table of contents), but I'm not sure what the calculation would be in the raster calculator to subtract it from the relevant band. If anyone knows, I'd appreciate some confirmation. Would it be as simple as "band"-[MINIMUM]?

Lucy Taylor I solved the problem using a Semi automatic classification plugin (SCP) in QGIS. It is automatic and straight forward in this software. Unlike ArcGIS, you don't need to use complex and manual maths (see link below). Just make sure you have all the necessary bands and the metadata file. If you are a maths whiz,  do it manually by feeding the required parameters in the DOS formula using the Raster calculator ...... each band has its min value since atmospheric effect is wavelength (band) dependent. Use histogram to explore the min and max values. The metadata file has most of the required parameters for DOS

https://media.readthedocs.org/pdf/semiautomaticclassificationmanual-v4/latest/semiautomaticclassificationmanual-v4.pdf 

Thanks so much, Makemie. I really appreciate your helpful response. I will explore QGIS and see how I go.

Dear colleagues, 

This can be easily done with  GRASS GIS by i.landsat.toar module.

And don't forget of Pat Chavez's suggestions.

ArcGis is not really suitable soft for satellite imagery analysis. 

http://grass.osgeo.org/

Hi, 

I am trying to perform an automatic way of process Landsat 8 from DN to surface reflectance. One of the 'pains' have been to get the kl value automatically, because I refused to use the minimun value of the whole scene.

So what I finally did was to get a DTM (I have to cut a bigger one with the extension of the current scene, because is not always the same) and run Fmask (https://code.google.com/p/fmask/) whit throws a very good water mask (and cloud, cloud-shadow, snow-ice, and clean terrain pixel).

Once I get these two rasters (Hillshade and Fmask) I masked every band getting only the values within the terrain shadows and water bodies, then I choose the minimun of these masked values as the kl value.

For me, this would be correct, but I'd appreciate it if someone could let me know its opinion.

Ps. The python code for this is here: https://github.com/LAST-EBD/Protocolo/tree/master/Metodos/Get_kl_csw