Has deconvolution been applied to nodes in a sensor network? Convolution, and blind deconvolution, is generally applicable where there is some "truth" function describing a physical process, and then some kind of distortion.
I have a bunch of sensors in a network that sense various parameters about the environment, and I'd like to see whether blind deconvolution could be used to deduce the most accurate point spread function.  
Has anyone had experience with this, and would recommend the application of deconvolution to a sensor network?  Thanks, much appreciated.
 A: 
Has deconvolution been applied to nodes in a sensor network?

Probably yes, maybe not, the question in the title is too broad.

I have a bunch of sensors in a network that sense various parameters about the environment, and I'd like to see whether blind deconvolution could be used to deduce the most accurate point spread function. 

You could always start with simulations and experiment what kind of noise levels the approach tolerates. If it does well in the simulations, and the results from a real environment appear to make sense too, you might have something there.

Has anyone had experience with this, and would recommend the application of deconvolution to a sensor network?

I can suggest something that I believe would be very novel in many fields of science regarding the use of deconvolution methods.
The convolution model is $h = f \star g$, where $h$ is an observed signal, $f$ is some source signal, and $g$ is the magic from the environment/point-spread function/filter depending on how you want to look at it. The $\star$ denotes convolution.
Now, it is known that convolution in the time-domain is equal to multiplication in the frequency-domain. That fact is known as the convolution theorem. So, suppose you transform the signals into frequency-domain. What happens for the mixing process? It becomes instantaneous, since you now multiply $f$ and $g$ instead of convolute.
So now what? If you have any instantaneous mixing model $\mathbf{x} = \mathbf{As}$ where only $\mathbf{x}$ is known, you can solve $\mathbf{A}$ and $\mathbf{s}$ with only few assumption (note that this applies to multivariate data too, as I indicate with the bold type face). The solution can be found with independent component analysis.
If you got interested, search for FastICA (an ICA algorithm developed by Aapo Hyvärinen) or InfomaxICA (an algorithm developed by Bell and Sejnowski). Don't miss a Matlab software package called 'Icasso', neither.
PS. If you do it I'd like to be a co-author :-)
