Confidence in pan-sharpened classification?

[Markenf]

I pan-sharpened one Landsat Image on ERDAS using the resolution merge resource, with brovey transform and nearest neighbor techniques. But I was told that in terms of land-use classification, it would be a mistake doing it, since pan-sharpening creates false pixels that would add erros. Is it true? Thank you.

[ahmad125]

I have this problem.

What is your erdas verion?

[Markenf]

It is Imagine 2013

[Arhanghelul]

Yes, is true. The errors will increase when you pansharp raster layers that have bigger resolution differences (ex: IKONOS, QuickBird and WV2 they have 1:4 spatial resolution differences). For the Landsat 7/8 the ratio is smaller, 1:2 (15 m panchromatic band, 30 m multispectral) so erorrs should be smaller. It also depends on what pansharpening method you use. Try different techniques.

 

If you want to classify a pansharped image, a better alternative solution to the traditional per-pixel methods - supervised and nonsupervised (Maximum Likelehood, ISODATA, etc) would be an OBIA classification. 

OBIA is less sensitive to false pixels because it first creates image objects (poligons) and then it classify these objects, rather than individual pixels. Theoretically, a pansharped image should produse better classification results when using OBIA.

 

eCognition and ENVI EX are the most common OBIA softwares.

 

Arhanghelul

Edited February 6, 2015 by Arhanghelul

[Markenf]

Thanks a lot!!

[mamadouba]

The easy answer is it all depends, which is also the difficult answer. It is not wrong to classify pan sharpened imagery, and there's plenty of peer-reviewed scientific journal articles that support this technique, but there are a few factors that need to be taken into account before applying this technique. First, there are many image fusion algorithms that are used to augment spatial context in the final product or to develop a more visually appealing "eye candy" picture. For example, pan sharpening can significantly change the original digital numbers depending on the technique. Ideally, if you have two pixels with the same digital number, the output values of the same pixels after performing an image transformation should be the same. However, this may not be the case and should serve as a warning. Second, what are the land cover/land use types that you wish to classify because this will affect the final map accuracy. A two category map, water and land, will always be more accurate than a 20 category map because of spectral confusion between class types. Third, does pansharpening actual contribute to the spatial scale of your analysis? if not, don't do it. And final, you don't know how accurate a final classified map is until you actually assess the accuracy. I'm assuming you are performing a supervised classification using ground truthing or a truth map. The only way to know if there is a difference in map accuracy between the native resolution image and the pansharpened product is if you apply the same training and validation data to each image dataset and compare the kappa or the Pontius et. al. confusion matrix statistic. If you are using unsupervised classification algorithms, all of the abovementioned points are irrelevant because you have no quantitative means to determine additive error or accuracy anyway.

[pasfans01]

depends on pan sharpening method you are using, there were some methods that relatively will not alter spectral properties of the multispectral bands when they are pan sharpened. Some of them are Ehlers FFT pan sharpen (implemented in ERDAS IMAGINE), Gram Schmidt Spectral Sharpening (implemented in ENVI), Yun Zhang UNB's Sharpening (implemented in PCI Geomatica), SFIM sharpening (implemented in ER Mapper), and Wavelet transform (Erdas Imagine, PCI Geomatica). 

[nadar]

I agree with you, on some satellites, the difference of resolution between pan and multispectral becomes ridiculous. Keep in mind that a ration 1:4 means you get 16 Pan pixels for one multispectral ! High contrast objects completely "smear" on the neighbours. I understand for marketing reasons, producers want to announce the highest figures, but....

I'm not even talking about the effective resolution of these images. If you want, you can resample a Landsat TM image to 10 cm !

 

I don't thin an algorithm will be able to reconstruct high res multispectral images suitable with classification. If the information is not there, you will never reconstruct it. On the same mood, you could also publish a movie captured with your cell phone on a Blu-Ray disc. It's feasible, but the results...

 

If your classification is mainly based on spectral properties (e.g. agriculture), I will rather work on pure spectral bands only.

If recognition is more based on shape and structure, Pan-sharpening may help.

I would rather use pan sharpened images for visual interpretation only.

[dr_s1ngh]

is it possible to do classification after pan sharpening? I think that it is wrong. One can do only visual interpretation of that image.

Edited May 20, 2015 by dr_s1ngh

[nadar]

I fully agree with dr_s1ngh

Edited May 21, 2015 by nadar

[Arhanghelul]

If it is ok to make a visual interpretation on a pansharpened image, than you can also apply an Object-based image classification (OBIA), no ?

OBIA does not classify each pixel, but it groups pixels with similar spectral/ color properties. OBIA is like visual interpretation, a good alternative to traditional per-pixel classifiers (Maximum Likelihood, ISODAT, etc).

Edited May 21, 2015 by Arhanghelul

[nadar]

If you use spectral properties to classify, they should be as pure as possible.

Visual interpretation using your brain is much more sophisticated than any computer-based classification, because it constantly switch between a local analysis (mainly texture and shape) and a global analysis (scene structure). Having a (big) computer slowly driving a car (in moderate traffic and on good roads with plenty of road signs) is one thing. Driving madly a junk car in Instambul traffic jams is another story. I'm not saying a computer is always more stupid than a turkish taxi driver, but as far as scene analysis is concerned, I will not put a penny on the machine...

[Arhanghelul]

If you use spectral properties to classify, they should be as pure as possible.

Visual interpretation using your brain is much more sophisticated than any computer-based classification, because it constantly switch between a local analysis (mainly texture and shape) and a global analysis (scene structure). Having a (big) computer slowly driving a car (in moderate traffic and on good roads with plenty of road signs) is one thing. Driving madly a junk car in Instambul traffic jams is another story. I'm not saying a computer is always more stupid than a turkish taxi driver, but as far as scene analysis is concerned, I will not put a penny on the machine...

 

When you have to make a multi-temporal analysis (like driving multiple cars at once on a road), you need to find an automatic or semi-automatic method to classify all the images as fast as possible.

OBIA can be used to create a hirarchical image segmentation (different scale levels) and a rule set to apply approximately to all images, obtaining satisfactory results.

This is a solution to visual image classification, when using pansharpened images.

OBIA is not "perfect", but is a good alternative.

 

This is my personal opinion, based on my experience.

 

Drive safe !

[pasfans01]

after all, its depends on accuracy assessment, if we could get a good classification result (validated thru accuracy assessment) from an spectrally distorted pan sharpening result, IMHO, it counts!, remote sensing is about discovery and exploration right? 

[Arhanghelul]

I agree.

 

[mamadouba]

You can 100% apply image classification to a pan-sharpened image.  There is plenty of peer-reviewed literature supporting this, including my own research.  Arhanghelul makes excellent points on OBIA and taking advantage of the textural and spectral information in the data, in addition to the variety of non-parametric classifiers that are capable of extracting meaningful information from fused datasets.  The final point from pasfans01 settles it entirely; if the accuracy assessment validates the results, then the approach is sound.

[sinan398]

Did you use the same type of images?