It depends on the dataset. No one can guarantee that SVM is better than ANN or ANN is better than DT. Setting parameters for each algorithm is also an important factor for improving classification accuracy. For most dataset, SVM and ANN provide higher classification accuracy than DT, but DT is the only classifier in these three which can identify important features from the model (the other two provide black-box model). You may try applying all these three classifiers to your data using WEKA and then compare the results.

It depends on the dataset. No one can guarantee that SVM is better than ANN or ANN is better than DT. Setting parameters for each algorithm is also an important factor for improving classification accuracy. For most dataset, SVM and ANN provide higher classification accuracy than DT, but DT is the only classifier in these three which can identify important features from the model (the other two provide black-box model). You may try applying all these three classifiers to your data using WEKA and then compare the results.

Thank you so much for your recommendation.

Deep convolution neural network (CNN) is a promising candidate since CNN outperforms state-of-the-art algorithms in image/object recognition, classification. However, CNN requires huge dataset for training 

CNN is very interesting. I will find more information about it. Thank you so much.

I agree with Ekarin that no absolute winner can be found in all cases. There are many papers around and SVM and ANN have become the people's favourite. If I can add something to the discussion, I think that more than the classifier itself it is important to find the correct representation to detect maize fields (= extracting the information that is really relevant). With a reduced, but good representation, even the classical LDA can do a great job. So have a look to DT or to SVMs that rank features (check MKL, for example) or even ANN with pruning. Or study the relations between features and classes and apply some smart feature selection beforehand. Just try not to take it as a blackbox, whatever the method you chose. Hope it helps.

I agree with all the previous comments. It may be wise to try several different algorithms to determine the most suitable candidate. Some other statistical methods such as maximum likelihood has been widely use but come at the cost of various assumptions made concerning the distribution of the data. Using maximum likelihood as an example the main assumption is that the data is multivariate normal in distribution. Other methods such as decision trees may not have this assumption but depending on the samples taken may create tree structures that are misleading if outliers are present. I guess to have a robust method I'd also suggest having a good sampling strategy for the data and to determine the distribution of class values. Methods such as transform divergence can work well in this respect.

Dear Chanika, nice to meet you here :). People gave you good suggestions. I would add about SAR imagery. You can try to use combination of the Amplitude Change detection and Coherent Change detection methods to extract agriculture areas. Good luck!

Dear Chanika 

Yes, I have an idea :) In machine learning, we can not say any method always ourperform the others. Because supervised learning or classification is based upon some criteria such as statistical distribution of training/testing data, resolution domain(spectral,spatial etc) of bands, the experience of the person who performs classification, the complexity of the selected geographic areas etc..... Moreover some classification algorithms such as SVM needs parameters to run and it needs optimum parameters to get highest performance. The optimum parameters are also based on the training/testing data and it's properties. Therefore we can not say any method always performs better than others....

And....

The classification accuracy could be changed based upon the fused data and it's algorithms...As an additional techniques, you can try ensemble classifiers since you would like to increas classification accuracy...Hope it helps...

Please see the article below. 

Björn Waske, Matthias Braun, Classifier ensembles for land cover mapping using multitemporal SAR imagery, ISPRS Journal of Photogrammetry and Remote Sensing, Volume 64, Issue 5, September 2009, Pages 450-457, ISSN 0924-2716, http://dx.doi.org/10.1016/j.isprsjprs.2009.01.003.

http://www.sciencedirect.com/science/article/pii/S0924271609000070