I have very high resolution images from LANDSAT 8 (5 out of 12 bands), which are of various administrative regions of a country. Each image is of variable dimensions, but generally of the order of [1500 X 1200 X 5].

My aim is to predict the population density from urban features visible on the images.

Since the number of images (and hence data points) is small, what is the best implementation strategy to build a model that can predict a value for population density based on these images?

  • $\begingroup$ You need to provide more detail. What do you want to predict from the images? $\endgroup$ Feb 24 '20 at 15:56
  • $\begingroup$ I am trying to predict population from the images. LANDSAT 8 has 30 m resolution bands and I believe that the model can map the Urban features to population estimate. I am worried about variable dimesnions and less number of images. $\endgroup$
    – anurag
    Feb 24 '20 at 15:58
  • 1
    $\begingroup$ Would it be an option to split the images into smaller segments (which could be of a fixed size) and use those for training? $\endgroup$ Feb 24 '20 at 16:54
  • $\begingroup$ My problem statement involves "looking" at the images as a whole - to learn spatially. I do not see how splitting will help! At the same time, any CNN based solution will effectively do the same. @OliverMason what you have suggested is available at rastervision.io but, as I said, splitting would be my last choice. $\endgroup$
    – anurag
    Feb 24 '20 at 18:11

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.