# What is the best machine learning technique to fuse two spatial data sets?

I have two data sets, containing points geometry (X,Y) and a recorded car exhaust parameter (let's say, RP value), of an area of interest (AOI). The datasets are spatially different, that is, the first data set is along side walk (X1, Y1, RP1) and the second data set (X2,Y2, RP2) is on the road center line (line split into equidistant 2 meters points).

The distance between the data along the side walk and the one on road center line is varying, at some locations, it is 3 - 6 meters and at some locations it is > 6 meters (let say, 6 - 20 meters range). This is due to the fact that this distance reflects varying road widths, lengths in a realistic, complex city landscape.

With the above data in hand, I want to fuse both data sets, considering the data along the side walk "more reliable" (thus higher weightage?), and compare the fused output with the reference data at limited locations in the AOI, to evaluate the data-fusion performance.

What is the best machine learning/data science technique to achieve the above? I am open to exploring several (or the "best candidate") technique(s) in Python, R, Matlab, for example. The focus for me is on the data fusion technique.

Ps. It is also possible to obtain information on road widths, lengths, building present or not, etc., if it is deemed "suitable" to include in the data processing.

• Would you like to have good position reliability? If yes, do you have the real positioning values in order to teach a neural network which values are the closest to the real ones? Jul 20 at 15:37
• Assuming real positioning values reflect distance between the points, then yes, i would like to have such reliability. It is possible to compute the distance for each point along the sidewalk to the points (or the nearest neighbor) on road center line. Jul 21 at 9:38

In terms of navigation, one of the most reliable algorithms is the Kalman Filter because it predicts directions according to previous points.

In your case, if you have 2 points measurement at each record, if you use a Kalman Filter, it would identify which one is the closest to the predicted value, without having outlines.

If you apply the Kalman Filter correctly, you would have a 3rd positioning value that would rectify the two others, and help you identify which one is the most reliable.

There are several libraries to achieve this:

1. PyKalman
2. FilterPy including experiments
3. Simdkalman

Be careful to set a good noise reduction value: too much noise reduction would make trajectories too precise and subject to outlines errors and too low noise reduction would make trajectories too blurred.

A good noise reduction value should be closer to the natural trajectory, which is smooth and clear.

• Thanks a lot, Nicolas! I will play Kalman filters and keep you posted. Jul 21 at 15:15