The purpose of my analysis is for examining the change trend in vegetation cover. I am using landsat images from early 1970s to most recent one in 2015.
Two widely used entirely image-based atmospheric corrections methods that I have developed are the 'Improved DOS' and 'COST' models that you may want to consider. If you google 'remote sensing chavez COST' you will see how others have used these techniques as well as compared them with other methods. Keep in mind that a very critical aspect of any entirely image based techniques is that a 'VALID DARK OBJECT' exist in the image for use as the minimum dark DN values. If a valid dark object does not exist the haze values will tend to be to large and the surface reflectance values will be over corrected and will under estimate the spectral reflectance values. Also, pay attention to the 'Improved DOS' method which suggest the use of realistic scattering models when computing the haze values for the various bands.
These methods are discussed in my 1988 and 1996 papers, plus additional information is contained in a 1989 paper.
For the purpose of vegetation change trends it might be enough and appropriate to use an image-based relative radiometric normalization method. Or to get absolute ground reflectance, use the absolute-normalization methos as defined in doi: 10.1016/j.rse.2006.03.008
These methods are relatively easy to carry by using general raster tools like map calculator and targeted just on the temporal change problems. You can find many publications on the topic with a search like this https://scholar.google.cz/scholar?q=relative+radiometric+normalization+remote+sensing&ie=UTF-8
I am currently working on development and practical implementation of spatially-variable variant of such method, and already have developed a tool for it in GRASS GIS. I am going to post an article on this hopefully soon, and release the tool as free software.
Edit: As there was some interest in it by other researchers, I have yesterday published preliminary version of the above mentioned tool. See https://www.researchgate.net/publication/272162808_i.grid.correl.atcor_0.9a?ev=prf_pub
Edit2: Since the first edit I uploaded several new releases, see https://www.researchgate.net/profile/Tomas_Brunclik/publications?pubType=dataset and find the latest one.
I agree with Chavz to use a DOS method for atmospheric correction. But I will recommend this method only if you dont have the ancillary data related to the atmospheric conditions during the image acquisitions.
As I have performed a comparitive study for different atmospheric correction methods using the in situ surface reflectance data (you can find the study link as following).
If you have the data for atmospheric conditions during image acquisitions, I will recommend to use 6S (a physical based atmospheric orrection method).
Best of luck.
Article Evaluation of atmospheric correction models and Landsat surf...
I read the your abstract and requested the full paper to get more detailed information about the comparison you did with the various correction procedures. In the mean time I am curious if you compare the straight forward DOS, the Improved DOS, or the COST models? Also, what time of year were the images collected? I would also be interested in what targets within the image were your minimum dark objects?
I have sent you the full text. The images were acquired during the months of January, October and December of 2013 coincident with the in situ data of surface reflectance collected using a hand held spectrometer. The images were having water and vegetation covers as dark targets.
Thank you for the additional information; a quick thought is that the images were collected when the sun elevation angle is quite low, and as I pointed out in the 1996 COST paper the TAUz (Cos(theta)) will tend to over correct and you have to be extra careful with correcting images collected in the winter time. Also, I am not sure if the water bodies are good 'valid' dark objects for haze correction, plus not sure about vegetation sites either. Did you attempt to find a dark shadowed area; if one does not exist then this could create a possible problem with the atmospheric corrections. Also, did you consider if the haze values represent realistic atmospheric scattering functions (see the 1988 paper).
Thank you for sending the full paper. After reading it I have a few comments:
* You compared the DOS method NOT the COST method published in the 1996 paper, even thou you reference the 1996 paper. The correction of the TAUz multiplicative parameter will make the predicted surface reflectance values larger, which should match your in-situ measured values better. Leaving this correction out will under estimate the values, as you saw for your mid to bright reflectance targets, but having haze values that are to large will also over correct and under estimate the values. If possible it would be good to get the haze DN values you selected and used for the correction, as well as the sun elevation angles of each image (mostly in December so relatively low sun angles).
* Did you check the haze values to see how well they fit a realistic scattering model as suggested in the 1988 paper you referenced? If you did, what was the power value of the scattering model you used?
Thanks for your comments on the paper, yes I have used the DOS method from the 1988 paper.
For selecting a DN haze value, I have selected the DN value from the histogram of Band1 where there was a sharp increase in the histogram. While for the scattering model, I have selected "Very Clear" e.g. the power of the scattering model was -4.
For example, for 7 December 2013 image used in the paper, I have selected the starting haze value from Band1 as 65 (DN haze = 65) and considered a very clear scattering model (Relative Scattering model = λ-4), thus the normalized DN haze values for the other multispectral bands (B2-B7) were as following;
Band2 20.4
Band3 15.1
Band4 7.2
Band5 6.0
Band7 4.2
Have I followed the method correctly, any further suggestions/comments will be appreciated.
Since this relates directly to your data set and is probably getting into more details than what others might be interested in I will follow up with you thru private messages. Sorry to the rest of you for taking this into so much details within this thread.
Now I am conducting atmospheric correction using FLAASH. Simultaneously, I used TriOS RAMSES in the field at the same time of Landsat archived. However, the results after atmospheric correction from Landsat are higher than TriOS. Could you tell me know how to calibrate TriOS to Landsat or Landsat to triOS. Or can I use your COST method instead FLAASH ? Actually, I read some researches and they shown that FLAASH is better than DOS. So, am I right ? If so, using FLAASH is fine ?
I need more information about what you are doing to be able to give you possible suggestions as to what is happening and what to try. If you answer the following questions I will be able to better understand the issues:
1. What do you mean that Landsat is higher than TriOS? Are the reflectance values higher for Landsat than those computed / measured in the field using TriOS?
2. Are you over vegetated or non-vegetated areas or are you over water ?
3. What are the sun elevation angles during the Landsat TM over flights / images?
4. Have you applied FLAASH, DOS, Improved DOS, and COST corrections to the Landsat data and compared them to the TriOS values to see which ones are computing values closer to TriOS?
5. One of the BIG issues when comparing satellite image data (like Landsat) to field spectral measurements is how well do the field measurements match a satellite image pixel. That is, were the field sampling done with sufficient accuracy to represent what the satellite image pixel is covering --- this is especially important when looking at the near-infrared portion of the spectrum and a mix of vegetation and soils exist within the satellite pixel. If the vegetation portion of the pixel is under sampled in the field the satellite reflectance values in the near-infrared bands will be higher (which could perhaps be the case in your study if you are in vegetated areas).
1. Yes, the Remote sensing reflectance values of Landsat are higher than TriOS RAMSES values. However, the trend is similar.
2. I am focusing on water surface to investigate the correlation between turbidity and remote sensing reflectance to develop the turbidity algorithm. You can see more detail the study area - Cam Ranh Bay, Vietnam on google earth.
3. I am using Landsat 8 OLi on Feb. 14 2016 to interpret the turbidity. Please take a look the metafile attached for more detail.
4. I use FLAASH and DOS. However, DOS seems not very appropriate and low accuracy. Improved DOS and COST no idea about them eventhough I read your papers already.
5. On the field, we measured 1 day difference. For turbidity as well as closed bay like Cam Rany bay, 1 day to 2 day difference does not a matter, I Think.
If you have email, could you please give me your email for further discussion.
I have done some field spectral measurements on water surfaces ranging from low to very heavy suspended sediment levels and turbidity including the Colorado River at the bottom of the Grand Canyon, San Francisco Bay, and coral reef in Hawaii. In these settings I would be concerned about the field data being 1 to 2 days different from the satellite image data. Also, a problem unique to water habitats is the possible glint, foam, and waves which can get averaged out in the relatively large satellite pixel footprint BUT not in the field spectral measurements. Also, the spectral reflectance values for water are quite low compared to on-land targets so an error of a couple of percent reflectance can be large. Since you are looking at dark reflectance values I am surprise the DOS method did not do better for you.
Below is a graph showing the relationship between suspended sediment levels and red to nir ratio of data collected on the Colorado River. Note that the nir band becomes more important as the sediment loads gets into the moderate to heavy; my guess is that within the bay you are studying the loads do not get nearly as high. My email address is [email protected]. Attachment did not work so will send by email later.
Just took a look at your study site (Cam Ranh Bay Vietnam) on google earth as you suggested and I have the following comments:
1. The waters look relatively 'clear' with very little turbidity or suspended sediment (at least in the ground photos attached to google earth). Getting information about turbidity at this relatively low levels is going to be difficult with remote sensors.
2. There is some wave action so this could impact the field measurements.
3. One of the problems with DOS (and other image based methods such as improved DOS and COST) is that I am not sure you have a 'valid' dark object to use as your haze value.
4. Looking at the information you sent on the spectral reflectance values of your sites shows that they are VERY low reflectance values. Have you computed the potential reflectance resolution of the Landsat TM DN values; you may be within the noise of the spectral resolution.
Thank for your kind information. It is helpful for me.
I used AAQ to measure turbidity in the field, and now I have to develop the relationship between turbidity and Rrs from Landsat or TriOS. I think the problem of myself is TriOS calibration.
I am reading your papers and try to calibrate TriOS as well.