Hi Adrian, 

A misinterpretation is something that you can always expect from image classifications. We cannot obtain a 100% accurate on the image interpretation results due to producer and user's error. Thus, what we can only do is to minimizing the error until a certain threshold, which people usually use > 0.8 for the limit of acceptable accuracy, even though it fully depends on the study purposes. In order to increase the accuracy, you can increase the number of training samples as well as use image composites or vegetation indices to assist the classification processes, as you have mentioned. In addition, you can always use more advanced classification methods like Object-based image analysis, neural network, etc. that based on the machine learning systems. A numerous study has been done using these advanced methods, which the results of these methods were much better than the conventional ones.

Hope it helps.

Regards,

Hero

I would recommend using rule-based classification by which you segregate vegetation types based on biophysical attributes, such as NDVI, soil, elevation, and rainfall rates.

If you use medium spatial resolution image, you might consider to classify with CART method

Hi Hero, thank you for your suggestions. I have heard of object-based image analysis (OBIA) but I'm not familiar with it. I'll definitely give it a try. Hopefully, it does help in improving the classification result.

Hi Rana, is rule-based classification the same as OBIA?

Hi Arif, could you please suggest some papers on CART.

Thank you.

I believe use the NDVI to separate vegetation from non-vegetation is a good strategy. Both object-based or rule-based methods are good methods. However, it also depends on the image type which you will use. The object-based method is more suitable for the high spatial resolution images, e.g., IKNOS, SPOT, etc. The rule based method is more applicable to the situation when other environmental information, such as soil, topography and landscape are used.

Hi Tao, I'm using Landsat 8 imagery of 30m spatial resolution. I've tried using OBIA as suggested by Hero but it still does not resolve the problem (I've done this with svm classifier of Orfeo Toolbox in QGIS). I suppose, as you've stated, OBIA would work better with high spatial resolution. Since I'm using only ms imagery for the classification, I don't know if rule-based method is applicable to my work. Thank you.

Hi Adrian, the key issue is not which method is the best, but what problem that you are facing. I suggest that you'd better check the classified image, and find the possible reasons of misclassification. So you may try to solve these problems. Best wishes.

To do LULC using Landsat 8 OLI, with a spatial resolution of 30 meters, it can not be obtained thorough results.

Using Landsat 8 OLI only resulted the map scale that is equivalent to a scale of 1: 100,000 (as recommended by Tobler). So it is not likely to obtain detail information.

To obtain LULC should not use vegetation indices, such as NDVI, because the spectral value will be simplified so that the classification would be more difficult

If the aim is only for distinguishing "vegetation" and "non-vegetation",  we can use NDVI.

I would rather suggest using Supervised Classification

Make training or sample as much as possible with the help of the FCC. Make sure that the different colors on the FCC has been sampled. Eg made of 50 samples

If we do not know the name of LULC when making sample, just use the code first.

Later these codes can be checked on the ground

In the final result, after the multi-spectral classification is complete, then some LULC classification results can be combined (merged) into one.

Eg "Rice field age of  2 weeks" with "Rice field age of 2 months" at the time of sampling should be separated. However, on the final result, the two can be combined

You can use NDVI and then classify it into classes. The forest cover should have the highest NDVI values (close to 1). Then compare the result visually to exmine the NDVI classes. When you are satisfied, mask the forest cover and then classify the other.

You may also use the  method proposed by Lloyd et al. (2016) which is called spectral variability vegetation index (SVVI). It is effectively distinguished between natural vegetation and agriculture area based on their results.

http://www.sciencedirect.com/science/article/pii/S0034425716302693

In their book Pattern Classification and Scene Analysis Duda and Hart said " We should have a healthy skepticism regarding studies that purport (claim) to demonstrate the overall superiority of a particular learning or recognition algorithm"

So errors and mis-classifications should be expected, however, you work on reducing them

Hi Bambang, prior to your response, I did try classification after NDVI, and as stated by you, it was not very effective. I was still not able to separate the different vegetation classes in my study site.

Hi Abdullah, using NDVI for my purpose has not helped. But I appreciate your recommendation. I will try SVVI, just out of curiosity. Thank you for the link.

Hi Abdelgadir, a very apt quote. That is what I have been trying to do.

Hi all, I have finally resolved this problem. I identified the areas having tea plantations on Google Earth and extracted them into shapefile. I used this shapefile to clip my ms imagery. I then classified the clipped image (having the tea plantations) and another image (without the tea plantations) separately. I then mosaicked the two classified images.

This proved to be the most effective way to do it.

Congrats

After having "fighting" with the problem, finally, you got the answer.

Now is my time to ask you to share your experience as to "HOW TO EXTRACT DATA FROM GOOGLE EARTH INTO SHAPEFILE"

Thanks  

Haha! It sure was a fight. Sorry, if I had misguided you. I meant I created a polygon of the area I was interested in, saved it as kml and then converted it to shapefile.

Thank you.

Good discovery Adrian. I want to believe this approach of extracting, clip and classify is practical when the "troublesome" feature is large block for easy digitizing the boundaries. This may not apply to features scatter across mosaic landscapes with varying patch size.

Perhaps you may want try object oriented classifier that considers other in classification other than the spectral values.    

https://pubs.usgs.gov/pp/1219/report.pdf

Find the leaf canopy reflectance of different vegetations in Landsat image. Based on that, you can classify it.

Hi Adrian, nice idea for your resolved way by extracted GE image to distinguish between tea and forest. Now, I want to hear your opinion, how if the image captured by drone which it is updated than GE's image