Problem
A mobile operator needs the signal level at any point of a city. Sales want to know where it is good to push SIM cards and routers. Network engineers want to know where it is bad to decide on a new base station or modernization. Drive tests cost a lot of time and man-hours. Crowdsourced measurements (Ookla Speedtest data bought by the operator, later phone reports under Minimization of Drive Tests, MDT) exist only where customers happen to be: dense along highways and near malls, empty elsewhere.
Example from the article: in some cities the Ookla dataset has one point in six months, and the model has nothing to work with there.
Solution in brief
Predict signal power (RSRP) at an unmeasured point with a weighted average of known neighbors, and choose the weights so that clusters do not dominate. Ordinary kriging does this with a covariance matrix of the points, and regression kriging adds a linear trend built from antenna, terrain, building and weather features. Validate by leave-one-out with all points within 75 m of the target removed.
Questions this case answers
- How to estimate a continuous quantity at unmeasured locations from sparse measurements that are clustered where the observers happen to be
- How to validate a spatial interpolation model honestly when the measured points are spatially clustered and unevenly dense
- How to interpolate sparse crowdsourced measurements into a continuous map
- How to validate a spatial interpolation model honestly when the data are clustered
Tools and methods
- KNN, IDW — baselines: plain neighbor average, then distance-weighted average
- Ordinary and regression (universal) kriging — weights from covariance; trend plus kriged residuals
- Local kriging — only the nearest points, to cut compute
- Jackknife with exclusion radius — validation for clustered data
- Ookla Speedtest data, OpenStreetMap, Meteostat — measurements, building features, precipitation
Google or a one-time email code. No password.