Hi David,

I think the problem here is not the classification method itself, but the fact that you are attempting a pixel-based classification, which doesn't deal well with SAR data due to speckle. Considering the high spatial resolution of stripmap data, the problem is even worse, because of the natural heterogeneity of the landscape. It would be just as difficult to classify an optical, high resolution image.

I think you should consider looking into image segmentation and object-based classification (of course, I'm biased, since I use it on my work). I don't know what access you have to commercial software, but there are some free alternatives. Take a look at https://sites.google.com/site/segsar2/home, for example.

Non-parametric classifiers are also a good suggestion, as radar data is seldom normally distributed. We have had good success with ALOS Palsar by extracting attributes from image objects created by segmentation (mean, standard deviation, etc.) as training samples, and running them through the WEKA data mining software (also free), using it to build a decision tree classifier. If you have a single image, the number of attributes is limited, but with two polarizations you should have at least mean HH and HV backscattering (or any other measure of central tendency), standard deviation of HH and HV backscattering (or any other measure of object heterogeneity), and polarization ratios. Once you build the tree on WEKA, you can take the values and reconstruct them in the software of your choice.

Integrating optical data is something that I always recommend, but at StripMap resolution I think you be very limited in terms of options, without sacrificing resolution.

All the best, and good luck!

Hi David, it would help to know what type of product you are using (ScanSAR, StripMap, etc.), what kind of landscape (urban, forest, mountainous, etc) you are looking at, and what classes you are trying to map. More detail on what you consider "strange" would also help. But from the start, I would't expect SAR results to resemble optical results, given the very different nature of each type of imagery. Even after speckle filtering, the heterogeneity of the SAR image and the limited number of bands will demand different approaches than what you would usually do for optical.

Your question is very unspecific: first, into which classes do you want to classify the data? Are the features in your feature space (hence, the HH, VV, ...) a good predictor for the classes you want to predict? Which predictor/model are you using? Which training data?

I just want to say that the strange results can be caused by many thinks: a bad model/predictor, wrong or bad training data, etc.

Please specify your set-up and in what term the results are "strange".

Dr. Silva and Herr. Senf, I am using StripMap mode in a vegetated terrain (generally flat) but with a variety of vegetation types within a wetland setting. As a first step I did Maximum Likelihood Classification using training set data developed from the imageries and confirmed by field visits and a complementary aerial flight campaign. I have not used SVM approach before. Might it give better results? You are proposing that I use other approaches, would you mind mentioning the ones that may work given the limited number of bands that I am working with? Thank you very much.

I am not that much into SAR, but from the classification side Maximum Likelihood (ML) should deal with a three-dimensional feature space. However, if the features are multicollinear, non-parametric (non-linear) methods such as SVM, RandomForests or Neural Networks could outperform ML, even though these methods are actualy made for high-dimensional problems. In generall, a three-dimensional feature space is very limited in predicting complex classes.

Beside some specific SAR methods, I suggest you trying following:

1. Compare the ML to SVM and other classifiers (try the EnMap Box: http://indus.caf.dlr.de/forum/). However, I think this will not change much.

2. Think about a fusion of remote sensing data. Hence, integrate other (optical) data.

3. Think about integrating other environmental data: climate, elevation, etc.

I think that the problem is that you simple do not have enough information to do a proper classification and therefore need a higher-dimensional, or new classification input.

Thank you Herr. Senf for the detailed response. I will definitely try all the ideas you have suggested and will update you on the progress I make with the datasets.

Hi David,

I think the problem here is not the classification method itself, but the fact that you are attempting a pixel-based classification, which doesn't deal well with SAR data due to speckle. Considering the high spatial resolution of stripmap data, the problem is even worse, because of the natural heterogeneity of the landscape. It would be just as difficult to classify an optical, high resolution image.

I think you should consider looking into image segmentation and object-based classification (of course, I'm biased, since I use it on my work). I don't know what access you have to commercial software, but there are some free alternatives. Take a look at https://sites.google.com/site/segsar2/home, for example.

Non-parametric classifiers are also a good suggestion, as radar data is seldom normally distributed. We have had good success with ALOS Palsar by extracting attributes from image objects created by segmentation (mean, standard deviation, etc.) as training samples, and running them through the WEKA data mining software (also free), using it to build a decision tree classifier. If you have a single image, the number of attributes is limited, but with two polarizations you should have at least mean HH and HV backscattering (or any other measure of central tendency), standard deviation of HH and HV backscattering (or any other measure of object heterogeneity), and polarization ratios. Once you build the tree on WEKA, you can take the values and reconstruct them in the software of your choice.

Integrating optical data is something that I always recommend, but at StripMap resolution I think you be very limited in terms of options, without sacrificing resolution.

All the best, and good luck!

I am sorry I am not able to give you the solution to your problem as i am not working with SAR data.

I have minimal experience with active sensor imagery, but I would think using texture to differentiate features would be useful (assuming what you're trying to detect has sufficiently different texture from its surroundings). Good luck, I'll be interested to hear how it goes. I've always felt that radar would be useful for regular monitoring of agriculture, however, I've never had the opportunity to investigate the ability to detect crop stress/disease. The ability to see through any cloud cover and also collect without sunlight are very desirable for monitoring rapidly changing land cover conditions.

Please correct me if I'm wrong, but doesn't the Cloude-Poitier decomposition (and others) only work with full polarimetric data, since it is based on decomposing the scattering matrix? As far as I know, TSX full polarimetric mode was experimental and very limited in availability, and it appears that the original poster is using Dual-pol data.

Thank you Richard, will take a look a the article, seems interesting!

Thanks Richard for the advice. I am working on PolSARPro. I will also look at the paper you have given.