1

I have a set (A) of geospatial raster data layers and a set (B) of point data. (B) includes numerical values that indicate the "suitability" of the specific site with a rising value. I know that (B) depends at least partly on geospatial information covered in (A).

Now my idea is to train an algorithm with my point data of (B) to find the patterns in (A) that contribute to the known suitability and then calculate for me in the areas which are not covered by point data a probability/possibility of suitability.

I thought either of a supervised neural network toolbox or fuzzy logic script to be applied in a GIS environment to do the operation.

Has anybody got an idea with which GIS-tool (either QGIS or ArcGIS) one could execute the analysis successfully and derive meaningful results?

  • 1
    The question you ask is pretty open-ended and broad, but certainly something that would be suitable for deep learning is some way or other. I would steer well clear of any desktop GIS and do this directly in Python, as this is where most research is focussed. Keras (on top of tensorflow) and Pytorch are currently flavour of the month. However, you are going to have do narrow you question considerably to get any sensible answers. – John Powell Apr 23 '19 at 9:17
  • 2
    Welcome to GIS SE. As a new user, please take the Tour, which explains how our "Focused question/Best answer" model operates. Unfortunately, broad questions which solicit opinions do not fit GIS SE. Certainly either tool could work with sufficient experience. In order for us to be useful you'll need to choose a path and begin work, asking focused questions as they occur. – Vince Apr 23 '19 at 9:24
  • Have you thought of fitting a generalised linear regression model before diving into machine learning systems? – Spacedman Apr 23 '19 at 10:31
  • Hello Spacedman I don't know if John Powell and Vince will be amused about our conversation but GLR might be an option as well, now that you mention it. However, having only point data for set (B) - and not raster data as for set (A) - the GLR would be limited to the pixels where point data is available I assume...to build the prediction function. – user140892 Apr 23 '19 at 11:54
  • Ah ha. The "point data" is a set of events. So model your points as a Poisson process with intensity dependent on your covariates. There's a few ways to do this, there's an implemention in R package lgcp, or you can build something with INLA. I think. Not an INLA expert myself... – Spacedman Apr 23 '19 at 12:51