Seperation of 40 classes in this small area of 5km2 where are you working? there is a lots of hetrogeniety and probably the RS classification needs much fine tuning.

Hi Deepak,

First of all, i believe that , separation of 40 different classes is not very suitable for your case. There are many different reasons (the amount of req. training data, ground truth generation, etc) . Therefore, merging these classes will be better -as Nicola said..

Moreover in order to achieve a good classification performance, you also have to consider other important topics:

-Dataset: the google images in your dataset should be obtained from the same satellite/aerial image . Otherwise, you may face with data-shift problem.

-Feature: The selected features must be discriminative and invariant: you can combine color and texture features.. ( GLCM, Gabor etc..)

Good Luck!

CS

note: Can you give the name of the studies proposed by "few researchers from China"

Thanks Caglar for your comment,

In my case I am familiar with the place I am working on, now as the aerial image is of very high resolution and different attributes can be easily distinguished, ground truthing and training data is really required I doubt.

I am not getting your point about data shift problem I would appreciate you if you can explain it for me.

Here I am attaching the paper "Exploring the Use of Google Earth Imagery and Object-Based Methods in Land UseCover Mapping" for your reference.

Dear Deepak,

are you aware about this publication? This may answer some of your queries:

"Google Earth and Virtual Visualizations in Geoscience Education and Research. " Edited by Steven J. Whitmeyer

https://rock.geosociety.org/Store/detail.aspx?id=SPE492

Are you aware about this publication which may answer some of your queries

"Google Earth and Virtual Visualizations in Geoscience Education and Research". Edited by Steven J. Whitmeyer

https://rock.geosociety.org/Store/detail.aspx?id=SPE492

Dear Deepak,

Data: The google earth image downloaded from google earth are acquired from various satellites (like IKONOS, QUICKBIRD and other high resolution imaging sensors). The specifications of these sensors vary and hence the data produced by them also varies. You can't simply use the multi-sensor data for analyses. The study you referring to from china used specifically the QUICKBIRD images.

Classification: The method you are adopting for classification is based on the spectral properties of the land surface object captured by the imaging sensors. With only three bands there is not enough spectral data to extract useful information from it i.e. the landuse/cover classification. The chinese study used an object based approach for classification that is very different from the method you are adopting (supervised and unsupervised classification).

Classification scheme: the level of classification you are trying to achieve should be in accordance with the study and the landuse/cover classes present in the study area.

Without good spectral resolution it will be very difficult to obtain good results. you told that you are using Google image, What is this? Original quick bird images or simply Google ? because if is the last you have an image that was several times transformed and in consecuence lost most of its spectral information . then it will not be usefull for your purposes.

a huge number of classes within small area? it seems that land parcels size is small. Especially for vegetation classification you need the NIR band because visible bands are strongly correlated.

I suggest to do a photo-interpretation, then import the KML file in a GIS environment.

The work is time-consuming but more accurate.

I believe that using only RGB bands for classifying satellite imagery would make you lose valuable information. And I I agree with some previous collegues that say 40 classes are way to many for RGB supervised classification. Maybe you could try object based classification, to focus on geometric features of your desired classes.

Deepak,

I guess you have now figured out the issue, with these above comments of the fellow researchers. What is your resolution of these “High resolution Google Earth Images”? If you’re using the inbuilt feature of Google Earth to download these, sometimes you still get pretty coarse images. Anyway since you’re using only RBG composite images (Which probably being highly manipulated in fitting to the Google earth) and lack of IR or other important bands, your accuracy will not be up to the expected level. Having these odds and you’re trying to pull the information from rather a small area will further add some error to the results. My suggestion is to go with fewer classes with larger AOI and compare the accuracy. Good luck with the study.

Dear Friend,

The spectral range of landforms are always unique. Use the high resolution image may provide good result when it have actual reflectance as it is object on the earth, As said Asato, google earth image may have reduced reflectance by sequence processes. Please check it first, I suggest you to calibrate DN value to reflectance conversion for individual bands before starting classification and also go for sub-pixel classification method. this may be increase accuracy

I agree with previous responses. However, you will probably achieve (slightly) better results if you use the RGB data to generate additional features. For instance, given that the imagery is of high resolution and asuming you exported it at its native (e.g. 0.5m) resolution, it might help running a principle component analysis (PCA) on the three bands and then to run a few texture measures on the first component. Then include these features as well as the PCA components (or maybe just the first component) in the unsupervised classification. Do not add more than five additional features as that might reduce the performance of your classifier (assuming you are using a parametric clustering algorithm such as isodata).

For such a small area, I would manually delineate and attribute the classes. Repeat with with 2-3 other interpreters then compare results.. For conflicting classes, either field check or make a expert decision.

40 classes, limited RGB bands calls for a spatial contextual analysis and the human brain is still better than a computer.

Thanks to all for their valuable comments.

Here 40 classes which I mentioned for supervised/Unsupervised classifications were supposed to be reduced into 5-6 classes after recoding i.e. aim was to make it into 5-6 classes like buildings, lawn, lake, road, bare soil etc.

Resolution of the image I am trying to classify is same as of images in Google earth, which is very fine as compare to SRTM or Landsat.

It seems absence of NIR band is a big problem for classification, I would try to explore object based approach.

Deepak,

Up to date, high resolution true colour 3D orthophoto imagery is better for landuse/landcover mapping however; reliability is only assured by systemmatic field checking using permanent benchmark transects; they can also serve for on-going monitoring purposes. When the mapping is complete and field checked, only then is it possible to employ false imagery lookup tables for calibrating spectral signatures for reliably identifying various landuse/land cover types.

After 40 years mapping and modelling using true and false imagery in a variety of GiS, and field checking work produced from multi-spectral false colour imagery, I am surprised there are still folk who fudge landuse/landcover mapping by ignoring or avoiding field checking protocols..... I have found a far better method is to learn and employ the habitat mapping methods that use unique spectral signatures to produce an infinite number of unique landuse/landcover units (not classes) using parametric codes to describe landuse/landcover attributes..... see Tane et al Auditing Reforested Loess Rangelands Ecological Indicators June 2014...its online already...Haikai

I agree with Haikai about the critical necessity of field checking. I have been classifying landsat images using an unsupervised and then a supervised classification, focussing on a 10 km by 10 km area that i am roughly familiar with. However, challenges arise in ascribing a mix of veg and soil to a 30 m pixel, and furthermore, differentiating between a fallow field and bare ground. This is hard to do both from spectral signatures as well as looking at a Google Earth image, unless one knows the place, which implies field checking. I guess how straightforward this is depends place to place upon the LC specifics. best of luck, and Deepak, this is an interesting question, very relevant for much of the world, thanks.

I think getting a high spatial resolution RGB Google Earth images results in a lower spectral resolution image which highly impacts the land cover mapping process, especially if your study area is large and the landscape heterogeneous. The most feasible solution in your case is to do the old fashion land cover delineation process which is manual digitizing. You have to have a standard mapping scale unit though and familiarity with image interpretation keys. This techniques might be time-consuming but worth it in your case and sometimes relying fully on automated classification methods can be misleading and requires post classification validations due to similarities in some features signatures. I have been adopting this technique in my current research due to the variability in land surface features in my study areas.

It will not be useful but only option is interpret and then save in KML to bring on the GE for better resolution.

You can get a good resolution RGB Images if you use the professional Google Earth take a many single clips and then make a mosaic of these images by using different softwares as GIS and Global Mapper.

I suggest you to use free Landsat images from USGS (Glovis / Earth-explorer) instead of Google earth and use Google earth as ground truth for classification of Landsat images. see my paper http://dx.doi.org/10.1016/j.sbspro.2011.07.044

even my area was more than 100+ksq

I suggest yo to use free Landsat images from USGS (Glovis or earthexpelore) instead of Google earth Images and using Google earth as ground Truth for classification.

see my paper: http://dx.doi.org/10.1016/j.sbspro.2011.07.044

If you are doing supervised classification as well, try to extract your ROIs for classes and subclasses from Google earth image and use these as training areas for a stacked layer of Landsat image, PCA first three components and any other product of image enhancement that can bring a contrast between the landuses you are interested in. I hope that you will be able to capture maximum variability. Later Google Earth image can be used for validating the output. Good Luck!

Could you give some information about your final classification system? How many classes? What are they? Land use or Land cover or hybrid?

If what you want is the classification of the Google image into different classes, then there are many ways to do it. For example, object-based classification or a simple segmentation are very useful to deal with high-resolution RGB images. They use contexture imformation as well as spectral information. If your goal is to get an accurate land use/land cover map, then using GE has potential problems. In order to produce a seamless coverage, GE intentionally distorted the images. So in many places, percentages, areas, and distances derived from GE are not accurate.

The better way is to interpret visually, but you may try high resolution segmentation and object oriented classification technique to achieve reasonable accuracy results.

I think object oriented classification technique would be a better choice.

I agree with all of the previous responses. The comments of Dr Stamou are important

Although you did not tell us the type of features you are trying to classify, whether urban, vegetation or other resources. This in the first instance will determine the type of images to use in terms of resolution. My suggestions are that you should first have a good base map which will give you some familiarities with the features you are trying to classify, you can then carry out an unsupervised classification followed by a supervised classification. I also recommend you use ILWIS software first to train yourself and then compare your result using other image processing software.

I agree with Asmat Ali, object oriented classification techniqueis a better choice. I've tried on an my area and gave very good results

I suggest use fusion process, Google earth image with others high resolution images like Quick bird or Spot images(IR band) you will increase your accuracy.

Since your study area is only 5 sq.km., visual digitization may be a good choice. Use S/W like Quantum GIS where you can digitize LULC classes using google map plugin.

You could go through this link its having answer which is related to your question. 

http://gis.stackexchange.com/questions/107524/image-classification-using-google-satellite-images

Hi, Deepak! Have you tried object-based method for the classification of GE images? Have you get a satisfying result?