Density kernel estimation

What's it all about?

A method is proposed for particle models to predict smoother concentration fields (and thus to speed them up accordingly), by assigning a density distribution to every particle, instead of the less efficient and (at the same time) less accurate box-counting method. When concentrations are to be estimated, the size of these density distributions (called kernels) is chosen by optimizing between a smooth spatial concentration density and its oversmoothing.

Kernels vs. box-counting

Stochastic particle models are the state-of-science method for modelling atmospheric dispersion. They simulate the released pollutant by a large number of particles. In most particle models the concentrations are estimated by counting the number of particles in a rectangular volume (box-counting). The effects of the choice of the width and of the position of these boxes on the estimated concentration is investigated. For the estimation of the concentration at a given point in space, it is shown that this numerical procedure can cause either oversmoothed predictions or too much scatter. As an alternative approach, the density kernel method to estimate concentrations is proposed, which optimizes bias and variance. It allows for a reduction of the number of particles simulated for the same accuracy. The efficiency of several density kernel shapes is compared, and methods for choosing their bandwidths are proposed. The relationship between the numerically motivated bandwidths and the description of the growth of a cluster of pollutant particles (puff dispersion) is discussed.

Further information...

...can be found in Chapter 4 of my PhD.
Or send me an e-mail:

Back to: Peter de Haan's home page. last update: 4 Oct 2000 - PdH