I am trying to apply machine learning to classify and distinguish crop types in my AOI, using optical satellite images. I don't have access to a pre-classified land use land cover reference map. And since this is a supervised classification, ground truth data is mandatory for a successful and accurate result.

So the next best solution is to organize a field campaign to collect data (GPS data, crop types, sewing dates, current state: growth stage of the plant...). I am not familiar with such a task, hence I wanted to get some insights on how to proceed since this is not an easy task and it is costly.

I already checked the previously asked questions here with similar tags, but I haven't found anything related or answers my question which is how to organize and prepare the field campaign. What I think is that I need first to visualize my AOI on Google Earth, check the land parcels with a good level of "greeness" and road accessibility, then check NDVI spectral signatures to somewhat infer the number of crops existing there and finally, choose locations (mark points) based on NDVI ensuring a somewhat equal number of locations for each possible crop type.

  • The process for setting this up would depend if you want to be able to get a probability that your results are correct. If you do, then you would need to randomly select points from the population of crop fields before going out. I would use Landsat rather than Google Earth so you can get the most recent imagery available (Landsat imagery is free)
    – jdavid05
    Commented Jun 22, 2020 at 16:40
  • 1
    Yes, I was planning on using the Sentinel 2 image (the closest to the field campaign date). However, I am not sure how to choose the points randomly in the first place. Should it be based on NDVI values (try to pinpoint locations that have different NDVI values because those will most likely represent different crop types?) because my goal is to choose the most accessible, and to cover as many possible existing crops as possible. Ofc with respect to a low cost.
    – Rim Sleimi
    Commented Jun 23, 2020 at 10:16
  • You might want to look at some other studies on this before starting your own. Crop surveys based on spectral analysis are common and they can give you some guidance. You only need to incorporate random sampling if you want the results to be scientifically valid. You could use strata sampling if you would like to make sure you get a sufficient amount of each crop. You could try to get the population of fields from supervised classification of the study area and then select a sample from that population randomly.
    – jdavid05
    Commented Jun 23, 2020 at 15:04
  • "You could try to get the population of fields from supervised classification of the study area and then select a sample from that population randomly." This is after the field survey right ?
    – Rim Sleimi
    Commented Jun 23, 2020 at 18:02

1 Answer 1


So basically, you need to use both Sentinel-2A image (the closest to the date of the field campaign) and google earth. Here are the steps on how to proceed:

  1. NDVI map and/or False color map is calculated based on the Sentinel-2A image.
  2. Create a mask for NDVI to keep only the values that are >=0.5 or >=0.6.
  3. The resulting parcels (after using the mask) are used to pinpoint (the same) parcels on google earth pro
  4. The resulting pinned points are exported as a kmz file and then to shapefile that can be manipulated using ArcMap.
  5. Usually going back and forth between the NDVI map and google earth pro is necessary to adjust the pinned points to the right parcels.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.