Background, in case it helps: My objective is to (eventually) fully automate the mass digitization of grain silos. To this end, I have set out to first find a way to automatically identify silos in a given raster image.

Question: Is there any way to automate the generation of training samples for the purpose of supervised classification in Arc, QGIS, or any other program?

I'm pawing through the literature and mainly finding things about convolutional neural networks and machine learning which, frankly, is above my pay grade. It's sadly looking like I could potentially be better off just with unsupervised classification, at the cost of accuracy and specificity.

As always, happy for any input or literature review recommendations!

  • With SVM (support vector machine) machine-learning algorithms you can achieve good classifications without much sample data . These are straightforward to run (e.g. using QGIS and Orfeo Toolbox), but the success will highly depend on the spectral and spatial resolution of your imagery – 15Step Jun 15 '18 at 22:06
  • What data and imagery are you working with? You would likely have very good success using a deep neural net with NAIP imagery, however, the learning curve is steep. – Aaron Jun 17 '18 at 23:57
  • I'm currently using NAIP, but have been investigating higher-res imagery as well, generally in the 15-30cm range. Beginning to suspect I might be out of my depth here, but I'm going to take today to continue my lit review and will report back with results! – instructorjack Jun 18 '18 at 13:52
  • I would look into using a neural net with a U-Net architecture. Here is an implementation on Kaggle that I've looked into in the past: kaggle.com/drn01z3/end-to-end-baseline-with-u-net-keras – Aaron Jun 19 '18 at 2:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.