We have a WSN (Wireless sensor network) deployed in an agricultural field for precision agriculture.
The sensors gathers some physical parameters (Temperatures, PH ...) essential for the automated treatment of this field (irrigation and so on).
We have a constraint in the number of sensors deployed as it doesn't make up all the surface ( let's say we have 20 sensors deployed in a 300x300m² surface)
When collecting data I want to be able to process it through some machine learning (or something else) so to transit from a discrete model (20 samples distributed in 300x300) to a continuous one (prediction over all the surface).
Do you have any idea how it could be done ? or is it even something doeable ?
PS1: This is not Time series forecasting, prediction of temperatures between dates of collections isn't the goal.
PS2: The data set I found Data package for automated in situ soil sensor network, provide the data of sensors in 42 instrumented locations in the field, so we can't use regular supervised regression methods, as we don't have data over all the surface to train our model.
So you want to interpolate between sensors. The simplest method, and there does not seem to be any reason to use anything more complicated, is "linear interepolation". Any point will be closest to 4 of the sensors. Let's say there are sensors at (a, a) which gives reading A, (a+b, a) which gives reading G, (a, a+b) which gives reading C, and (a+b, a+b) (b is the distance between sensors) which gives reading D. Now suppose the point (x,y) lies within those four points. The distance from (x,y) to (a, a) is $\sqrt{(x-a)^2+ (y-a)^2}= D_1$. The distance from (x,y) to (a+b, a) is $\sqrt{(x- a-b)^2+ (y-a)^2}= D_2$. The distance from (x,y) to (a, a+b) is $\sqrt{(x-a)^2+ (y-a-b)^2}= D_3$. And the distance from (x,y) to (a+b,a+b) is $\sqrt{(x-a-b)^2+ (y-a-b)^2}= D_4$.
The "weighted average" at (x,y) is $\frac{AD_1+ BD_2+ CD_3+ DD_4}{D_1+ D_2+ D_3+ D_4}$.