Algorithm 1. STEM.
Figure 1. Quadtree.
Left: eBird data locations showing the varying density of observations. Center: Two realizations of quadtree generated stixels, red and blue.
Right: Average quadtree stixel size (in degrees) follows the density of observations. (Color version of figure presented in electronic version
of AI Magazine).
that can be modeled by the mixture. In STEM λ is a
fixed, universal parameter that does not vary with
location. It can be estimated through cross-validation
to identify the scale of analysis best supported by
AdaSTEM proposes an adaptive scheme based on
tree data structures (Samet 2006) where stixel size λ(s)
varies with location s as a function of data density
(Fink, Damoulas, and Dave 2013). Letting the stixel
size λ vary with data density allows the mixture bet-
ter to exploit unevenly distributed data in the pres-
ence of a multiscale signal. In densely sampled
regions λ will be small and the base models can adapt
to fine-scale signals producing low bias estimators. In
sparsely sampled regions λ will be large and base
models are forced to adapt to large-scale signals pro-
ducing low variance estimators.
The center panel of figure 1 shows two partitions