I am performing supervised classification of Sentinel 2 imagery using a Random Forest Classifier. I wish to select the training data from the image. Could anyone please tell me the method to efficiently perform this?
Hi Shivam,
A similar questions are answered in the following links. I hope it's useful.
https://www.researchgate.net/post/How_can_I_choose_a_training_set_from_database
https://www.researchgate.net/post/What_is_the_best_way_to_divide_a_dataset_into_training_and_test_sets
And also you can find more answers here:
https://stackoverflow.com/questions/17200114/how-to-split-data-into-training-testing-sets-using-sample-function
Best regards,
Moslem
I would recommend to formulate your question in a bit more specific way. There are numerous aspects about selecting training data for a supervised classification in an adequate way. For example some important aspects are:
1. The training data should be reliable - that is if you intend to collect them directly in the Sentinel 2 images, you might want to search for some additional high-resolution imagery where you can actually identify the exact land-cover class visually and reliably. One option for this is loading the Sentinel-2 images into QGIS (the open-source GIS environment) and additionally load GoogleEarth high resolution imagery via the openlayers plug-in for QGIS (can be added to QGIS using the Plugin-menu of QGIS)
2. The training data include all important land-cover classes within the satellite scene to avoid wrong allocations of pixels to your classes of interest. Identifying all relevant and spectrally separable classes is not always an easy task. In some cases it can be interesting to conduct an unsupervised classification of the image with variable class numbers to identify how many land-cover classes can be reliably separated based on the spectral information available (you can for example use the isodata classifier implemented in EAS SNAP toolbox - which you probably also use for pre-processing the Sentinel-2 data).
3. Everything related to sampling and validation. This relates to an adequate number of training samples, the distribution of training samples in the images, the applied validation strategy (cross-validation, iterative validation with data-splits, etc.). There is some classic literature on this topic available. See for example the works of Congalton. Furthermore, I can highly recommend the book of Max Kuhn on "Applied predictive Modelling" where different validation and resampling strategies as well as their pros and cons are explained in a very helpful way.
As I said, it might be helpful for you to refine your question so that it is more clear to colleagues what you are interested in.
Hope that helps.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images