in my current dataset I have 30 years of max daily temperature history for the whole USA area (points on a lon/lat grid, each point is 10 km apart). I would like to predict if a max daily temperature will be at least 25 degrees. So my dataset looks like:

lon | lat | year | day_of_year | event_25_degree
1.11| 2.22| 1980 | 334         | 0
1.11| 2.22| 1980 | 335         | 0
1.34| 2.56| 2020 | 111         | 1

I'm really new to spatial analysis and I'm not sure what model should I use. My first guess was to use some type of splines. I think that the effect of day_of_year and lon, lat variables should be nonlinear so the model should look something like:

event ~ year + s(day_of_year) + te(lon, lat)

where s is spline smoother and te is tensor product smoother. On the other hand I was also reading about usages of kriging and Markov Random Fields (for example in bamlss package: http://www.bamlss.org/). My questions is what model would be best suited for this kind of task ? Should I avoid any one of mentioned methods when modeling large area (like whole USA) or modeling/predicting into the far future (like a year ahead) ?

  • $\begingroup$ How the data were measured is important. I strongly doubt your grid consists of summaries of actual observations, because there is no such grid of meteorological stations and it is implausible that satellites or other remote sensors have been available continuously for 40 years to measure daily temperatures. Thus, the spatial statistical structure of your data almost certainly depends (strongly) on the procedures used to create this grid of data. The first thing for you to research, then, is the metadata on how these data were created and processed. $\endgroup$
    – whuber
    Sep 27, 2021 at 13:09


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.