Another aspect that is typically disregarded is the probability of belonging to a given class. Usually classifiers provide a map with the class each pixel belongs to, but many of them can also provide the probability with which that pixel belongs to each class. For some pixels the probability between two classes is very similar (e.g. 35% and 34%), but the assigned class is the highest one, while in fact it could as well be the other class too. So the confidence of the classification of this pixel is not so high. It can help you discriminate which pixels to trust and which not.

This information is available, but seldom used.

If you want to approach change detection in classifying a case study, definitely supervised-based methods will be more appropriate. In most cases, unsupervised methods are not pretty reliable, they just can be used to bring out some useful regions of interest. You should investigate which features and methods (SVM, Random Forests, ANN) will be more suitable for your case.

Beyond doubt the supervised classification will give you a better result if you have preparered your training sites nicely and but unsupervised should be good enough if the landcover is not very hetrogenous and thsi will take much less effort and will be quick.

Even with supervise classification you must be careful, since there are land cover types such as agricultural that evolve in time due to phenology of vegetation, while the class has not changed. Or, depending on your spatial resolution, you might find static objects with vertical structure that change shading in different times of the year, even if the images are acquired at the same time of the day. All these, and any other you might think of, are susceptible to be included in the training.

In fact, the different methods available will depend on the characteristics of your problem: resolution, revisit, number of images, period of time,... which you have not specified.

It really depends on the resolution of the examined images. If you have high resolution imagery such as IKONOS WV-2 etc, supervised classification is far better than unsupervised. If you work with low resolution data, then unsupervised could be more convenient and easy.

That's depend on scale & accuracy u want , i prefer use unsupervised first to show me more explain after that i can capture significant table to use supervised

undoubtedly, It is depend on ground control points which are available. If you have enough GCPs, supervised method is the best. You know that unsupervised method is just used when satellite image is low in spatial resolution or you have not enough information about probable class numbers. it will give you preliminary information to select classification level.

Behnam you mean the ground truth points because ground control points usually can improve the positional accuracy only it has nothing to do with spectral characteristics.

Kapa coefficient and general accuracy coefficient will improve when GCPs are enough in the number and accurate in sampling.

For any change detection the simplest method is that first you need to classify the images of Time 1 and Time 2 and then run the change detection algorithm in any software like ERDAS. The accuracy of the classified images (T1 AND T2) will govern the accuracy of the change detection.

You can directly go for the the change detection between two images but ERRORS may increase if the two images are acquired in two different illumination condition and if the relief of the area is higher.

Therefore better classify the image using any algorithm and then do the change detection. Supervised Maximum likelihood classification can do the job.

Thanks all for your useful recommendations

Behrooz which supervised classiffication are you using. I would suggest run an unsupervised and see how many classes are your getting, adjust thresholds by ISODATA and see, may be if it fits your expectation use it, else use a supervized classification. Also when you do the two make a comparision between the two classifications.

Dear Saif Uddin

thank you so much for your opinion .please read my article in below link.

https://www.researchgate.net/publication/259640276_An_approach_for_land_cover_classification_system_by_using_NDVI_data_in_arid_and_semiarid_region

Article An approach for land cover classification system by using ND...

Hello Behrooz,

First i would like to say that your area of interest is in NDVI data in arid and semiarid region and it is very good work. so from my point of you it should come under target enhancement at sub pixel level, so we require high resolution image and after completing soft classification it is better to do for subpixel mapping(i.e., Super Resolution Mapping) which gives high accuracy and information extraction is also better with this, ok thank you.

Behrooz yours is an interesting article. I read it. However as per my experience on west of Persian Gulf NDVI is slightly less suited for classification in Arid areas for teh reason taht vegetation density is extremely low and often seasonal.

Dear Suresh Merugu

thank you so much

please answer me where can i download this data

We are offering an excellent forum for this type of discussions. Why not attending ForestSat 2014? Deadline for abstracts on April 30th, visit http://forestsat2014.com

supervised classification requires training data sets to perform the classification and unsupervised training data sets not used. supervised classification provides better result. But unsupervised as a advantage of identifying the distinct spectral class presents in image. my view is u can perform both and compare the result.

Dear all

thank you so much for you kindly attention

what is scale and resolution role in the select of classification methods.

It changes the classes that you can identify. With lower spatial resolutions urban areas are just a blur, and a single class could suffice, whereas at higher resolutions you could separate it into roofs, roads, dirt roads, etc. The other way around also happens, with coarser resolutions showing superstructures that are not evident at higher ones and prone to be classified.

With respect to spectral resolution a larger number of bands allows discriminating a larger number of classes and subclasses.

At this point it is important to note that "class" is an abstract concept and it is subjective, either to the operator under supervised classification; or to the algorithm under unsupervised classification.

For the same image, two operators and two algorithms will produce different sets of classes.

In the same manner, the same scene with different resolutions (either spectral and spatial) will result in a different set of classes even under the same operator, or the same algorithm.

At the end, you have to test on a representative subset of you data the performances of different methods for your particular case, and chose the one that provides the results that best suit your objective.

Another aspect that is typically disregarded is the probability of belonging to a given class. Usually classifiers provide a map with the class each pixel belongs to, but many of them can also provide the probability with which that pixel belongs to each class. For some pixels the probability between two classes is very similar (e.g. 35% and 34%), but the assigned class is the highest one, while in fact it could as well be the other class too. So the confidence of the classification of this pixel is not so high. It can help you discriminate which pixels to trust and which not.

This information is available, but seldom used.

Dear  Luis Alonso

thank you so much . how can i get this information?