Since SNR is affected by factors not present in graph topology, incorporating CNN for map image of area can be helpful. Basically we would 'convolve' parts of image per edge using GPS of node endpoints.
The top part represents the graph (processed via GNN) and bottom part represents the image of the map (processed via CNN).
This is really hard to implement because we would have to factor in the path edge follows on map. Easier (but still hard) to accomplish with for loops, but infeasible as the dataset size increases. Harder to accomplish with vectors, but much faster.
Since SNR is affected by factors not present in graph topology, incorporating CNN for map image of area can be helpful. Basically we would 'convolve' parts of image per edge using GPS of node endpoints.
The top part represents the graph (processed via GNN) and bottom part represents the image of the map (processed via CNN).
This is really hard to implement because we would have to factor in the path edge follows on map. Easier (but still hard) to accomplish with for loops, but infeasible as the dataset size increases. Harder to accomplish with vectors, but much faster.