How to handle multi-channel 2-D geo-spatial grid like data samples in machine learning with number of features associated with each grid?

0

I am looking into a problem wherein the whole geographic area is divided into number of bins/pixels so we get nxn matrix covering whole region. Now each bin/pixel has number of parameters/features associated with it e.g., number of buildings in that bin, number of people living in that bin, poverty level of that bin, crime level in that bin, etc. This whole information represents one sample of the training dataset i.e., we have this matrix like data for different geographic regions with corresponding labels. How to best handle this kind of dataset for machine learning task e.g., ML trained on some nxn grids for different areas like this will classify labels for some unseen test nxn grids. I am thinking that in term of CNNs, it may be represented in terms of channels so each channel represents associated features of a bin. Whats your suggestion?

One sample of dataset

Dataset

hasanfarooq

Posted 2020-08-18T00:10:17.900

Reputation: 11

Using multiple channels is the usual way of dealing with data like yours with several measurements per spatial location. Is there a particular reason you think it may not be appropriate for your task? – bogovicj – 2020-08-18T11:49:40.590

Thanks! Actually there are hundreds of measurements for different spatial location so number of channels will be quite large. Secondly these measurements are available for specific type of bins only i.e., if bin has specific type than these measurements are there otherwise not relevant. I just wanted to know if some other way is more appropriate for these type of datasets. – hasanfarooq – 2020-08-18T17:34:38.960

If the number of features/channels becomes an issue I'd recommend doing some dimensionality reduction (e.g. PCA or something fancier if necessary) to find correlations between your measurements and reduce your channel count. What is "appropriate" entirely depends on the assumptions and tradeoffs you're willing to make. – bogovicj – 2020-08-19T12:42:14.200

No answers