I am also looking for this solution

I am assuming that you have certain hyper-spectral images which you want to analyze but they are corrupted by the shadows cast on the object of interest. Links to some papers that you may find helpful are given below. One of the papers is about image in-painting for cloud removal and employs a special class of wavelets called bandlets. Although this may not be specific to your problem but I think wavelet based techniques may be helpful in shadow removal.

http://proceedings.spiedigitallibrary.org/proceeding.aspx?articleid=868906

http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6050213&tag=1

http://proceedings.spiedigitallibrary.org/proceeding.aspx?articleid=836433

http://www.sciencedirect.com/science/article/pii/S0924271613000415

http://www.sciencedirect.com/science/article/pii/S0034425709000807

http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4804804

http://www.sciencedirect.com/science/article/pii/S0034425709001199

Dear Mohammad,

Thank you for the links. I will certainly examine this information.

Dear Matthew,

I'm doing both: classification and various radiometric analyses. The latter is most critical  as classification and mapping of plant communities can be fulfiled by other images acquired in approprate conditions but investigations of plant cover dynamics must include as much data as possible.

Thank you for the link.

As for classification, you should keep in mind that by landsat images having 30 or 60m spatial resolution we cannot see exactly shadows arise from plant cover so we cannot tell apart plants and its shadows. We just fix decrease of pixel reflectance values and when there is no reference image to compare it comes out a problem. Sometimes plant cover fall into the same class with water bodies...

If the resolution is 30 to 60 m I think that then you will have to look for some characteristics of the shadows e.g. whether they dominate different wavelengths or even if they can be differentiated based on the some probabilistic measures. I have heard of a new technique in learning though it is quite new and may not have found many applications as such but you may be able to use it. It tries to find intrinsic geometry of different classes of objects and has been used effectively in case of small datasets. I am not sure if it will help you or not but I send the link if you like

Yes, Mohammad, send me a link, please. I'd like to explore every available technique

You probably work in high latitude areas with dominance of few species and large intraannual illumination variability, under conditions I a'm unfamiliar with.

Yet, let me suggest you consider exploring principal components analysis using Landsat time series to try to detect areas with significant intraannual and interannual variations, in particular, for the purpose of acquiring sample areas for classification. Of course, PCA results are highly dependent on the stats of your time series and seldom allow to make generalizations for systematic work over several different areas/timeseries. They also may be difficult to correlate to physical magnitudes to develop reliable numerical empirical relationships. Yet, they can help identifying features/transitions relevant for both cover classification and cover dynamics at least on an ad-hoc, case-specific basis (in case you are unfamiliar with PCA, most Image Processing textbooks offer reasonable reviews of the method).

Let me also suggest that you look at work using spectral mixture models. A recent reference related to those models that I have at hand may be:

https://www.researchgate.net/publication/272523428_Relationships_between_dominant_plant_species_fractional_cover_and_Land_Surface_Temperature_in_a_Mediterranean_ecosystem

Article Relationships between dominant plant species, fractional cov...

Thank you for an answer and link! It seems to be interesting. And if

PCA can be applied in this case, I'm going to try this. I'm familiar with the method of course.

I'm working with images of north Caspian sea region...

Dear Maria,

   I suggest you look for some kind of relation between the spectral values of the target points in your data in a form of a ratio to generate new features/variables.

As you may certainly know it, ratioing is one of the effective methods to eliminate the effect of shadow and slope (as it is used in NDVI).

Regards

Peyman

you can check the Image processing software tool ( GRAM++)developed by IIT Bombay.

Dear Peyman,

Unfortunately, only rationing doesn't eliminate shadow effect in my case... 

Dear Sandip

Thank you for the reference but can you please describe by what algorithms this soft corrects low sun elevation? What are the benefits of the program you refer in terms of a problem mentioned in my question?

Dear Maria,

I suggest you a number of method to solve your shadow problem that:

1) Use a topographic correction method, you can use atcor 2 toolbox for erdas or you can find some tools for ArcGIS.

2) Use vegetation indices to reduce shadow effect such as NDVI as a layer in your study. Because image wavebands effecting from the shadow in same way so ratio images can see also shadowed places. Not only use NDVI but also use the GVI, SAVI and NDWI for vegetation analyses.

Dear Onur,

Unfortunately these indices also work not well enough with low elevation angles. I guess it should be any kind of preprocessing transformation.... radiometric normalization maybe... and i'm going to try application of TCT directly to these images

Dear Maria,

  I did not mean "only rationing" as for simply calculating ratio between the two bands. Life is never that easy.

I don't think TCT or PCA can be of a great help to you since they are mainly a type of dimension reduction methods. They do not affect the quality of your incoming data.

Here are two publications you may decide to take a look at them. One is about tuning the system to detect certain type of cloud and one may use it to tune it on other things.

The other article is about optimizing NDVI to better fit the requirements.

"Optimizing Cloud Detection Index for MODIS Imagery Using Genetic Algorithm"

https://www.researchgate.net/publication/224944583_Optimizing_Cloud_Detection_Index_for_MODIS_Imagery_Using_Genetic_Algorithm

"NDVI Optimization Using Genetic Algorithm"

https://www.researchgate.net/publication/261194640_NDVI_Optimization_Using_Genetic_Algorithm

I may have misunderstood your question or my answers may not be so helpful, but I just wanted to be of some help and I wish you luck in your work.

Regards

Peyman

Conference Paper Optimizing Cloud Detection Index for MODIS Imagery Using Gen...

Conference Paper NDVI Optimization Using Genetic Algorithm

Dear Peyman,

Thank you very much for your detailed suggestions and links. Now I see what exactly you mean! It is most probably the subject i'm looking for. It's most likely, I misunderstood your first answer being tired yesterday. Sorry. 

I'm really grateful for your support.

And great thanks to all researchers who expressed the opinion on the subject I asked for.

Yours, Sincerely

Maria

Lots of good discussions on what may or may not help.  I will add to the discussion with a few comments:

* If you are going to use mostly (perhaps just) Landsat TM with 30m resolution for vegetation (ag fields?) analyses then very few pixels will be in complete shadow, and in fact, the percent of the pixel covered by shadow will vary through out the image and mostly less than 50%.  Therefore, you will be dealing with a 'mix pixel' type of problem.  As you know the amount of shadowing will be dependent on several factors, including the time of year (low sun angle problem) and the spacing and height of the vegetation.

* An interesting experiment might be to start with a very high-resolution image (say 2m) which will contain pixels that are in 'full shadow' and use it as input to generate digitally compressed resolution images of 5m, 10m, 15m, 20m, and 30m.  Look at the spectral characteristic of the vegetation of interest in the different spatial resolution images which will contain different amounts of shadowing within the  mix pixels.  Do this with a high sun angle image and with a low sun angle image; this will give you some information as to how the vegetation spectral characteristics at your study site will vary as a function of sun angle because of shadowing caused by the vegetation.

* With my background in radiometric calibration (Improved DOS and COST) I would say that you have to look at the calibration of the temporal images (including atmospheric haze corrections), however, beware that methods like the two I mention will tend to over correct images with low sun angles (lower than say 35 degrees).  In this case, perhaps what might work best is to do absolute radiometric correction of a high sun angle image and then use relative radiometric correction to slave the low sun angle images onto the high sun angle radiometric master.

* When you say you are looking at 'plant cover dynamics' are you talking about temporal or spatial dynamics ?   For temporal dynamics 30m resolution may be acceptable but for spatial dynamics my guess is that images with at least 5m or better resolution might be more useful.

This is probably longer than most readers are interested in so I will stop for now.

Pat

Dear Pat,

Thanks a lot for such detailed comments!

I don't know what about others, but as for me your post seems not that long.

As you mentioned in your first * topic I answer : yes, I'm going to use mostly TM data but not only TM. Though if I can succeed in processing at least of TM data I will be quite happy.

About your second topic: I began to thought towards the same things when I recieved the first answers to my question dealt with high resolution images...but I was so hoping that anybody have already done such experiments...on other territories.

So thank you for providing detailed instructions on what I can do. Along with Peyman's and some other suggestions this make me hope I can finally solve a problem.

As for the 3rd topic:

Did you meant a RRN method (Relative Radiometric Normalisation), I mentioned above?

And at the end answering your questions from the forth topic I've to say that I'm talking about both: temporal and spatial dynamics. I agree that 30 (or 60m in some cases) may be not enough for studying of spatial dynamics, but the area of my interest is very specific. Vegetation seasons may vary greatly in length and also vegetation activity as well as areas occupied with various plant communities (mostly grasslands). I've never faced before such extreme variations... Such a way as for my own experience I can suggest that in river deltas Landsat TM, ETM+ and OLI provide optimal ratio of spatial resolution and revisiting interval to monitor plant cover in such specific area.

Thank you for the help!

Your issue is analogous to low and/or shaded slopes in rugged terrain. You might browse this paper of mine for a simplistic, relative normalization approach.

http://www.geocarto.com.hk/cgi-bin/pages1/june00/topographic.pdf

There are also a variety of more precise radiometric correction techniques. I haven't worked on this in several years now, but Landsat 8 with its high radiometric resolution would seem to have great opportunity to improve the technique. I would match the correction/normalization to your goal (simple/relative for classification or more complex/radiometric if you are trying to derive changes or diagnostics from veg canopy, cover or other biophysical indices.)