you must to be at least 30 GCP for each land use type. of course, it depend on to location of GCP. you must check the statistical parameters of each GCP

GCPs are generally used for image registration. The numbers are decided on the basis of RMSE (

Bagher Ghermezcheshmeh Dharmendra Singh Thank you both for your answers. Very helpful!

@Prakhar Rawal

There is NO definite number of training samples to serve as input in classification stage, however it is essential that training data is representative.

Let me elaborate,

Make sure that spectral variability of each 'information class' to be classified is taken into account. For instance, if information class of Water is found to have uniform spectral response characteristics over the region of interest, then only one training area polygon would be enough to represent the class. On the contrary, if there is spectral heterogeneity within water class (e.g. turbid, clear etc.) you will have to delineate training areas, accordingly.

Actually, ultimate objective is to prepare statistics to describe the spectral response pattern for each landcover type to be classified. Therefore, it should be noticed that all information classes are adequately represented in training statistics for optimal classification results.

I hope it is helpful.

GCP= Ground Control Points use for image georeference. But I think You ask for Ground truth data that support image classification. You need to collect two types of data. One for image analysis and the other for accuracy assessment. Actually, as I know there is no definition for the number of ground truth locations. But it is better to collect points that are uniformly distributed over your entire area.

Thanks Janaki Sandamali

it depends on the area of the image, pixel size and the number of classes. usually for each class you need at least 15 samples. of course you must check the statistics parameters of samples data such as mean and SD for each class.