I have done land cover classification in ARCGIS (supervised raster classification using multiple landsat bands) and I have had some success in sub-dividing my classification regions up, then collapsing them back together afterwards. I work on braiding sand bed rivers where the exposed sand has different reflectance depending on its degree of water logging, and water has different reflectance based on its sediment load (and its depth to a degree). Rather than trying to generate a single class for "water", I created a series of supervised classes for different "types" of water. Then once I'd done this, took my resulting classification raster and collapsed all my "waters" down into a single water class (and the same for sand, vegetation, etc).

In your case it might be worth, rather than trying to treat your shadowed and unshadowed areas of the same type as one thing, split them (mentally/methodologically) at the shadow line and then try and do a supervised classification on "shadow water" and "non shadow water", then put them together into one class afterwards.

I can't think of another way around this, as fundamentally you are trying to do something based on the "colour" (in a crude sense) of the data, and areas which you know are the same thing are actually completely different colours. There is no way you can get a classification that will automatically put them as the same, as they have different reflectance. I'm not aware of any program/technique which can strip out colour differences caused by shadows, as that is the only other (theoretical) approach that would work.

i am not so sure how helpful this links will be. but i hope they are worth having a look. 

https://www.google.co.in/url?sa=t&rct=j&q=&esrc=s&source=web&cd=1&cad=rja&uact=8&ved=0ahUKEwjJwMmhmJnUAhUKtI8KHbV7BkEQFggqMAA&url=http%3A%2F%2Fwww.mdpi.com%2F2072-4292%2F7%2F10%2F13507%2Fpdf&usg=AFQjCNHowQi5UJpmL7BWlD7i3kVopAFPpA&sig2=SB2TswEo17Ic_Y-x-tXSvQ

https://www.google.co.in/url?sa=t&rct=j&q=&esrc=s&source=web&cd=2&cad=rja&uact=8&ved=0ahUKEwjJwMmhmJnUAhUKtI8KHbV7BkEQFgg0MAE&url=http%3A%2F%2Fwww.mdpi.com%2F2073-4441%2F7%2F2%2F794%2Fpdf&usg=AFQjCNF-aw2X4HLS3GVQjWYB-Y8Ki8-cRg&sig2=atuDTLA3An5joOjYt45TXQ

https://www.google.co.in/url?sa=t&rct=j&q=&esrc=s&source=web&cd=3&cad=rja&uact=8&ved=0ahUKEwjJwMmhmJnUAhUKtI8KHbV7BkEQFgg-MAI&url=https%3A%2F%2Fearthobservatory.nasa.gov%2FFeatures%2FColorImage%2Fpage2.php&usg=AFQjCNG9Xjm-IrB5XL1wolxPQ4ihYAzrXQ&sig2=leB0zMhvrOZmeWxBK3L8zQ

https://www.google.co.in/url?sa=t&rct=j&q=&esrc=s&source=web&cd=5&cad=rja&uact=8&ved=0ahUKEwjJwMmhmJnUAhUKtI8KHbV7BkEQFghGMAQ&url=https%3A%2F%2Flandsat.usgs.gov%2Fdifferences-pre-collection-cc-algorithms-and-cf-mask&usg=AFQjCNHRfgR762dOhaUMBaREZ1rLekXywA&sig2=yEVVByVTKUxJy8clPsv4SQ

https://www.google.co.in/url?sa=t&rct=j&q=&esrc=s&source=web&cd=6&cad=rja&uact=8&ved=0ahUKEwjJwMmhmJnUAhUKtI8KHbV7BkEQFghMMAU&url=https%3A%2F%2Flandsat.usgs.gov%2Flandsat-surface-reflectance-quality-assessment&usg=AFQjCNEZCG6zVdHoA8IKn6T7hNUehj-QCQ&sig2=u4Mwi37vhrKJ1cYjUpfXJA

I have done land cover classification in ARCGIS (supervised raster classification using multiple landsat bands) and I have had some success in sub-dividing my classification regions up, then collapsing them back together afterwards. I work on braiding sand bed rivers where the exposed sand has different reflectance depending on its degree of water logging, and water has different reflectance based on its sediment load (and its depth to a degree). Rather than trying to generate a single class for "water", I created a series of supervised classes for different "types" of water. Then once I'd done this, took my resulting classification raster and collapsed all my "waters" down into a single water class (and the same for sand, vegetation, etc).

In your case it might be worth, rather than trying to treat your shadowed and unshadowed areas of the same type as one thing, split them (mentally/methodologically) at the shadow line and then try and do a supervised classification on "shadow water" and "non shadow water", then put them together into one class afterwards.

I can't think of another way around this, as fundamentally you are trying to do something based on the "colour" (in a crude sense) of the data, and areas which you know are the same thing are actually completely different colours. There is no way you can get a classification that will automatically put them as the same, as they have different reflectance. I'm not aware of any program/technique which can strip out colour differences caused by shadows, as that is the only other (theoretical) approach that would work.

Dear Mohan,

did you try with SAR imagery ? I don't know if it is a good path ...

https://en.wikipedia.org/wiki/Synthetic_aperture_radar

You may find some methodologics orientations in these articles :

In English :

http://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Lin_Deep_LAC_Deep_2015_CVPR_paper.html

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4299033/

Courses and books for SAR imagery :

http://www.ldeo.columbia.edu/res/fac/rsvlab/fundamentals_e.pdf

http://www.jstor.org/stable/40510993?seq=1#page_scan_tab_contents

In french :

http://www.lthe.fr/PagePerso/gratiot/publi/Baghdadi_etal_05SFPT.pdf

For the acquisition :

http://eoedu.belspo.be/fr/liens/galeries.asp?section=7

It is difficult to differentiate, but the location is already indicative, and the presence of clouds or mountainous regions nearby. In the case of lakes, it is expected to find the value 0 for the pixel, in case this does not occur it is suggested that if an atmospheric correction

The Shuttle Radar Topography Mission produced a Surface Water Data Base that may be of use. Lakes and larger rivers are included, but of course it's just a snap-shot as of 2000. Links are available at:  https://en.wikipedia.org/wiki/SRTM_Water_Body_Data 

-Tom Farr, Jet Propulsion Laboratory

Good suggestions have already been made and I wanted to just add one more possible procedure you may want to consider to see if it could help:

If you generate RATIO images of spectral bands that are next to each other the

topographic shadow affect will be similar and to a first order the shadows will

be removed.  Perhaps using the ratio images as input to your classification may help; under ideal conditions you might be able to just do a simple threshold on the ratio

images to get a good first draft of what you are attempting to generate.

Good luck with your work,

Pat

You can use a digital elevation model (SRTM, A LOS, GDEM, ASTER, for example) and calculate the slope, aspect and roughness for example. A decision tree between the elevation products and the different indexes can solve your problem.

You may build an algorithm to do this for you. Try to read the following article to have a better idea. 

Santoro, F., Tarantino, E., Figorito, B., Gualano, S., D'Onghia, A.M., 2011. A tree counting algorithm for precision agriculture tasks. International Journal of Digital Earth, 1-9.

See the following links:

http://vault.sierraclub.org/john_muir_exhibit/writings/the_mountains_of_california/chapter_6.aspx

These looks promising:

https://nau.edu/uploadedFiles/Academic/CEFNS/NatSci/SESES/Research/Geology/Roop_2009.pdf

https://www.wta.org/news/magazine/magazine/viewing-the-handiwork-of-glaciers

http://www.usu.edu/geo/luminlab/cwyshnytzky_Thesis.pdf

Good luck!

If that is a clear imagery and orthorectified one approach is to use OBIA image analysis to classify them if there are other features other than true color RGB can also used as input for a multidimensional segmentation. You can use ENVI to preprocess the images ans then use eCognition for the classification work and use back ENVi for some post processing.

Please use visual imagery with a simultaneously taken infrared imagery. To compare contours and wherever contours of both visual and infrared show contour similarities we can assume absence of Shadows  and ice deposited areas give brighter appearance . But small pieces of ice deposits may pause problem.

My answer can be composed of 3 parts: 

1) WATER. Lake recognition is equivalent to water extraction. Because sediment under lake has little influence on differentiating water from other features, unless some algae is floating in the lake.

2) ICE. You can use the visible bands (blue,green,red) to extract water from ice. Because the reflectance of ice is much higher than water in visible band. Alternatively, the difference between visible bands and infrared bands can also be used.

3) SHADOW. Shadow spectrum is extremely similar to water spectrum. They are very difficult or impossible to be divided if considering spectral characters only. There are two kinds of shadows: cloud shadow and hill shadow.

3.1) Hill Shadow. Scholars often use other auxiliary data to remove hill shadow, such as DEM. Because places covered by hill shadow are on the hillside, where the slope is distributed heterogeneously. This heterogeneity can be expressed by DEM gradient (Yamazaki et al., 2015). Object with low reflectance and smaller DEM gradient is water, with low reflectance and higher DEM gradient is hill shadow.

3.2) Cloud Shadow. It is more difficult to distinguish water from cloud shadow. There are two methods, but they both need an amount of work. One is using time series data (Mueller et al., 2016), based on the fact that cloud location is various in different time and water location is more stable. But this method can ignore extreme events like flood. Another is using specific cloud shadow screening algorithm (Zhu, et al., 2012). This method uses the geometrical relationship between cloud height, sunlight and shadow to recognize cloud and cloud shadow.

You can refer to the attachment I make to understand the spectral difference of different objects.

References:

Yamazaki, D., et al. 2015.Development of a global ~90mwater bodymap singmulti-temporal Landsat images. Remote Sensing of Environment.

Mueller, et al. 2016. Water observations fromspace:Mapping surfacewater from25 years of Landsat imagery across Australia. Remote Sensing of Environment.

Zhu, Z,, et al. 2012. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sensing of Environment.

Supervised classification with well known traing site.  Training site may be chosen by Dgps or GPS coordinates(according to spatial resolution). Use thermal infrared image to differentiate the temperature between ice and water.  NDVI also can try same feature present  in shadow portion .all the best. 

It may better to use SAR images additionally  like Sentinel-1. SAR images are sensitive to water

You can remove shadow using topographic normalization. Then use infrared bands to separate water and some other band ratio to separate ice. Regarding topographic normalization may be this paper can help you.

https://www.researchgate.net/publication/303824782_The_effect_of_topographic_normalization_on_fractional_tree_cover_mapping_in_tropical_mountains_An_assessment_based_on_seasonal_Landsat_time_series

Best regards

Hari

Article The effect of topographic normalization on fractional tree c...

SAR image can give better idea about soil Moisture. No idea about ice or snow. 

When we use Infrared imagery to compare contour symmetry we know the shadows do alter very little on temperature and temperature contours. And shadows and clouds move to show their interactions. Due to angle of incidence of visual light on objects cause problems in using colour spectra

You can separate them with Land Surface Temperature 

If you are working with Sentinel2 and Landsat8 images you have these possibilities:

a) Perform a false color that differentiate them. The SWIR/NIR/green combination should work in these cases. In the case of the Sentinel 2, band12/band8/band3 and in the case of the Landsat 8: band7/band5/band3. This false color highlight snow in light blue and water in deep blue.

b) Other possibility is to make use of the high-level products of Sentinel 2 that offer some direct mask like snow, water, cloud shadows, etc.

c) Finally, a supervised classification with good trainning areas could be another solution in this regard.

You can try the following procedure:

1. See the metadata of the satellite images and note down the sun elevation angle.

2. Take a DEM and generate a hill shade image using the sun elevation angle (May use ERDAS or ArcGIS software).

3. The hill shade image consist of bright and dark pixels. Dark pixels corresponds to the topographic shadow at the time of imaging. This additional information from Hill shade image will help you to eliminate almost all topographic shadow areas in the satellite images.

I have tried this for detection of landslides in hills using OBIA. You may also refer my following paper.

MARTHA, T.R., KERLE, N., JETTEN, V., VAN WESTEN, C.J. and VINOD KUMAR, K. 2010. Characterising spectral, spatial and morphometric properties of landslides for automatic detection using object-oriented methods. Geomorphology 116, 24-36.

http://www.sciencedirect.com/science/article/pii/S0169555X09004292

Have you tried a simple supervised classification, to see whether the lake surfaces are actually within the same reflectance range as the shadow/land/ice/snow in all bands? That might be a logical first step, before jumping into ratios.

Dear Mohan,

In addition to all the suggestions made by colleagues, especially the use of DEM, I would like to suggest additional option, i.e. use of calibrated/optimized NDVI.

You may select different materials in your scenery and optimize the NDVI in such a way that it detects your target material with more precision.

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 would like to remind you that ratioing feature in NDVI by its nature will help you with dealing with shadow and slops in the scenery.

I think there is similar discussion in RG that you might be interested in:

https://www.researchgate.net/post/Is_there_any_reliable_techniques_for_remote_sensed_data_processing_on_images_obtained_with_low_sun_elevation_angle?_tpcectx=qa_overview_following&_trid=DaqnxyHDi4D7dz8TUbkMykSD_

Good luck

Peyman

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

Conference Paper NDVI Optimization Using Genetic Algorithm

Dear Mohan Bahadur Chand,

Unfortunately your question is not clear for me. It seems that it does not fit my expertise.

Best regards

Márton Veress

You could use a model of elevation to generate the area of drainage and confluence and hence delimit what would be ice or cloud and shadow

You may also want to try Object Based Image Analysis (OBIA) instead of pixel based classification. With OBIA in eCognition Software, you can differentiate objects based on spectral signature, geometry as well as context. In short, you can resolve/classify any object which you can visually detect with your eyes on any image. It involves machine learning so you have to be prepared to learn some basic coding and scripting. Good luck. 

Dear Mohan Bahadur Chand, you could do the surface temperature analysis, then classify lower temperature area, which is the ice area.

IF you use NDVI index you will get a good results

Literature there are plenty of paper to detect and remove shadow from images. One of the proposed methods is to work in a different color space, e.g. the YCbCr. Coupling different color levels is quite easy to detect areas in shade.

Hi Mohan, I hope you're well!

The suggestions above are very good places to start, this is a tough problem.  Some additional thoughts based on my experiences (see https://www.cambridge.org/core/journals/journal-of-glaciology/article/spatial-seasonal-and-interannual-variability-of-supraglacial-ponds-in-the-langtang-valley-of-nepal-19992013/5DC43C5E42FC87B19855C9993152E21A/core-reader)- 

1) Projecting shadows across terrain is a worthwhile start (some studies make a correction for deep shadows, even, but you need good radiometric resolution to do this, which OLI and S2 both have but TM/ETM+ do not).  The choice of DEM for projecting shadows is crucial, however - problems with both SRTM and ASTER DEMs in the Langtang region meant that projected shadows were inadequate.

2) The surface-type segmentation difficulty is due to the simultaneous illumination differences (shadow or not), material differences (wet or not), and phase confusion (liquid or frozen water).  This is too much for a single metric to worth with. Although I was able to get NDWI/NDSI/and band-ratios to work (note that this also required atmospheric correction so that ratios were in comparable ranges!), more sophisticated spectral analyses (e.g. spectral unmixing, or even just PCA) would be more effective at distinguishing between the surface types you mention - these methods take into account more spectral information at once.

3) Even with those successes, note that there s a significant bias in identifiable features to larger-sized ponds due to the sensor resolutions (here, S2 is a big step up from Landsat).  Some studies have gone on to only classify groups of 4 pixels or more! Take a look at Watson and others (2016) to see the impact in terms of % of ponded area identifiable...

4) Seasonality is another issue to be careful of - isolated ponds grow and shrink seasonally, and may drain/flood annually as part of the broader hydrological system of the glacier. 

5) Classification with surface temperature is another great idea, but is also limited by the MUCH coarser resolution of thermal bands (60m/120m).

I'm happy to discuss further with you via email or otherwise, but identifying your specific research question will be very important - it may be more suitable to use either different imagery or to select specific scenes rather than to analyse many images.

All the best,

Evan

Thank you all for your great suggestions. Many of you suggested very good idea, Hopefully, I can apply those techniques to distinguish the very small water bodies without effect of shadows, ice and snow pixels. As I am working to detect very small ponds, so resolution is most important, and therefore looking for still high resolution image besides Sentinel 2. 

Yes, Semon, that is very good idea to take a different samples from different water bodies, either with shadow or without shadow. Somehow, I tried it already, not successfully, I want try it more to get better one.

Herve, I read about SAR imagery but no idea how to use them to detect water bodies or which software use for analysis, if you suggest some idea, I would be happy to go through them and learn it.

Cary, In my case usually water bodies supposed to have small number of pixel than shadow due to tall mountains and more pixel with shadows. Yes, there is choice to remove shadow with greater number of pixels. I should use the DEM to remove the shadow part, where slope very high and less potentiality for pond formation. 

Tom, for me I think those data are not useful as I am looking for very small feature and also most of ponds are forming after the 2000 in most cases. Thank you very much for your suggestions.

Thank you all

Chavez, I really want to do the band ratio that are next to each other, let me try. Thank you very much for suggestion. Also, Many people are grateful to you as you made the DOS atmospheric correction freely available in open source software. It is very useful. Thank you.

Bernardino,  Thank you very much for your suggestion, use of Ecognition software is difficult for me for the OBIA as it is expensive, otherwise, it could help me.

Dear Kai, you really made clear to me. I am going to use DEM for the shadow. But sometimes, it is still difficulty to distinguish in which I already applied it. I will go through those papers suggested by you.

Dear Hari Dai, thank you very much for suggestion, I am going to use topographic correction, If I am not wrong is it a same as topographic normalization.

Gordana, Yes, I will use land surface temperature.

Thank you all

I am sorry. I don't have any answer at the moment.

I think you have to know the spectral reflectance of a shadow, snow and ice along the electromagnetic spectrum and you then use the results for classification.

Since areas with those 3 classes (shadow, ice and snow) have different reflectance values (DN Values), you can take samples as points for each class and measure the DN values. You can get a range of DN values for each class, for example shadow; 50-120, ice; 130-180 etc. use these classes in raster calculator in ArcGis or QGis to split your image accordingly.

Try with supervised raster classification in ARCGIS 

Sorry to reply late... 

Please it's simple. Apply Linear Spectral Mixture Analysis and the job is done. This can be done smartly by using Idrisi Software. You can also read about the Ridd's V-I-s Model. Thanks and I hope it bails you out. 

Yes, Susana, I will try for it and already tried it, but still to go through high level Sentinel 2 data.

Tapas, you are right, we can use the digital elevation model as reference to know about slope and hillshade, which is useful.

Jill, the main problem is even water bodies has different reflectance due to the amount of sediment it has, most of times when use the threshold like NDWI, MIDWI, the value for water, shadow and ice is not so different, so could not differentiate the water bodies.

Peyman, I calculated the NDVI but still have to learn optimization. It is difficult form NDVI to separate the water bodies, as far as I understand.

Niccolo, I did not understand YCbCr, could you give some more ideas about it.

Thank you all for your valuable inputs.

Dear Evan,

Thank you very much for suggesting very good points. I am following your paper for my current study. My first target is use the Sentinel 2 images with 10 m resolution for seasonal variation and then for the long term variation, I will use the Landsat data. Of course, I need to preprocess the data for consistency, now some atmospheric correction tool i.e. DOS, DOS1, 6S are available freely. I also tried to topographic correction using tool available in QGIS but could not get good result, I am trying on it. If I could better result, I think it can help in detecting shadows. Regarding the project the shadow using DEM seems a little bit difficult for me, I could not understand, could you explain it for me. Yes, Evan, there are many issues need to consider, but it is very difficult to get very high resolution images though I am trying to find out, but that I can use only for validation or referencing. I will be happy to hear from you more.

Thank you,

Mohan, I'm not sure if this question is still open. There are indeed many valuable suggestions so far. Just wanted to add that, you may want to look at the Quality Assesment (QA) band that is included with the "Level-1 GeoTIFF Data Product” of Landsat that you download from USGS. QA band might be helpful to identify shadows when overlayed with a True Color Composite e.g. of the same Landsat scene. Although I must say I did not find it very accurate over high relief areas such as Himalayas as compared to the plains. What are the coordinates of this supra-glacial lake?

https://landsat.usgs.gov/collectionqualityband