Hi,

It is difficult to say which one is better because it depends on the distribution of data (different conditions, different measurements).

The advantage of SVM is global optimal and multi-dimensional data, while Ensemble is that it doesn't require hyper-parameter tunning. (It is good to do both methods under the same distribution of data and compare their results)

However, I suppose that satellite images are much variant. So you should make features more abstract to make data more invariant using:

• AutoEncoder:

  Article Deep Learning-Based Classification of Hyperspectral Data

• CNN:

  Article Deep Convolutional Neural Networks for Hyperspectral Image C...

• or augment data using GAN:

  Article Generative Adversarial Networks-Based Semi-Supervised Learni...

Best regards,

Dear Phan,

Your question is to classify Landsat or Sentinel to obtain a land use land cover map. My answer depends on the details you want in your final map (i.e. includes types of forests, crops,...etc) and on the available training and verification data. In general I would try all mentioned supervised algorithms SVM, Random Forests, Deep Forest and even Artificial Neural Network to get different results. I would also use unsupervised clustering algorithms such as K-means, ISODATA, Fuzzy C-Means, SOM. What is left to do is to verify your results using your field work data. Although, it is time consuming mission, but it will give you confidence in your work and experience.

Thank Taha Sadeghi Chorsi , M. Awad and Hai Huynh for your helpful answers.

Actually, I do know many factors (e.g., types of land cover, level of landscape complexity and diversity, quantity and quality of training and testing data, the numbers of categories) affect the accuracy of a LULC map. What I am concerning is the most suitable classification method for a general type LULC map. Anyway, maybe each has its own advantages and disadvantages.

I suggested that uses SVM, because it get a high deficiency in classification problems specialties if your Data are linear separated

Yes, it is really a factor and matter of confusion to chose particular method and algorithm. actually, the accuracy of of classification is depend on total area covered under classification. For small areal extent, signature based supervised algorithms SVM can offers best result. But for large areal extent, unsupervised clustering algorithms such ISODATA can be considered. In this regards, I would like to suggest to use SVM if your area of interest is small other wise you can try with unsupervised clustering algorithms.

Some links are attached here, hope you may get help;

http://dataaspirant.com/2017/01/13/support-vector-machine-algorithm/

https://www.isprs.org/proceedings/XXXV/congress/comm4/papers/505.pdf

Follow the doi - DOI:10.5120/21088-3779

Thanks Phan Cao Duong