0
$\begingroup$

I am new to ML and have some experience with building CNN models. I recently got involved with a research project and here is the task I have to work on:

I've been given some (latitudes,longitudes) points with cellular signal strength(float numbers) on those particular points. And I've been given the location, (lat,longs) of cellular towers in a particular area. Also I've been given the terrain data, i.e. the (latitude,longitude) locations of areas where there are buildings, trees, hills etc. that can cause disruption of signals.

I have to predict the signal strengths for new areas given all this data(terrain data and tower sites). Basically, the signal strength will depend on the nearby areas(i.e. if they are hills or buildings or trees and so on) and how close a tower is to a particular point.

I've been trying to find a model for this for a week now with no apparent progress. The problem is that every model I've seen(CNNs, Variational Autoencoders, GANs) take in the image data and produce image data. What I have to do is predict numerical values of signal strengths on a map using the image(terrain) data. Any ideas which model can be used?

$\endgroup$
1
$\begingroup$

I would try a CNN architecture like U-net with following input planes :
- one plane for signal strength with negative value for unknown value
- N planes to encode the terrain type (with one-hot encoding)
- one plane to encode towers (the float value can indicate the power of the antenna and 0 for no tower)
All input planes have same dimensions than the area of interest.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Can you clarify a little bit more please? By one plane for signal strength, do you mean a 1D Convolution layer? Also, how do I tell the model that "on this image of map, this is the magnitude of signal strength here.", i.e. how do I account for the magnitude of signal strengths on the image? $\endgroup$ – Divyansh Jun 26 '19 at 7:20
  • $\begingroup$ By plane I mean 2D matrix of numbers. For example a RGB image has 3 planes, one for red, one for green and one for blue intensities. Regarding your second question, you don't need to tell what the planes represent/encode, the model will try to predict what is expected (thanks to ground truth plane and loss) whatever you give it in input. $\endgroup$ – Ismael EL ATIFI Jun 26 '19 at 20:38

Your Answer

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

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