------ Filtering: (applied to any except MB TE BA or BO)
5 point binomial filter (ie coeffs 1 4 6 4 1) with some tricks (over-weighting
available data) to avoid distortions due to gaps and at ends of profiles. To
maintain integrity of results, where gap corrections are made the depth is
correspondingly corrected.
 

------- Interp to 2m: (only applied to MB TE BA or BO)
Interpolate to 2m in the range 0 to max depth of cast, using a 4 point
quadratic fit. Maximum gap size limits increase with depth.
 

------- Binning:
In the first 3 cases, binning is applied to generate binned profiles from
which stats are generated
-- Boxcar: for each cast, a simple average is made of all data in each depth
  (or temperature) bin. The fastest algorithms cannot handle inversions in
  the binning variable. Mine will but it does require constant bin size!
 
-- Subsample: for each cast, the nearest value (within the bin only) to the
   the bin centre is used for each bin. Fastest method.

-- Central filter: a locally-weighted mean is evaluated for each bin of each
   cast. The weights start a fairly constant fall off close to the bin centre
   and the lengthscale of the filter gives 93% of the weight inside the bin
   and approx zero weight at 1 binlength beyond the bin. A total weight
   threshold means that a value will not be returned for a bin centred on a
   data gap of more than 1.3 binlengths.
   - weight function  1/(1+(dist/lscl)^4)      lscl=binlength/3
 
   Speed is between boxcar and sub-sample, and is faster with constant binsize

-- All Values: no binning of individual profiles - stats done on ALL
   observations in a given bin. This means that high res casts have much more
   influence on stats than low res (esp. if latter are not interp to 2m).
   This is also the slowest method.
 

--------- The stats:
The stats calculated are:
# mean    (1 value [for given depth for given time period] required to calc)
# standard deviation            (6 values required to calc)
# 1 2.5 97.5 & 99th percentiles (200 values required to calc)

Time periods are:
# no separate time periods if < 60 profiles
# yr/4, from start of year, if  60-120 profiles
# yr/6, from start of year, if  120-200 profiles
# monthly if >200 profiles
 

-- t0, t100, t250:  simply finds the nearests values to the given depths, if
   within 5.5m of that depth.

-- MLD1: interpolated depths of T(ref) +/- 0.5, where T(ref) is first value
  between 5 and 10m. Note can be triggered by a T inversion! Does not look
  below 20m gaps. All sorts of issues, like what to do when water mixed to
  bottom of profile (eg inshore). Just return a value where test succeeds, for
  now.

-- MLD2 non-interpolated depth (ie depth of observation) where gradient first
  exceeds +/- 0.015 C/m below 9m. If true at first obs below 9m, work upwards
  until it fails. Idea is that we want to avoid near-surface stuff transients
  unless is clearly a very shallow ML, in which case find just how shallow it
  is. Don't look below 10m gaps. Again issues of profiles mixed to bottom.

-- t(z). Simple stats using one of the binning methods.
 

-- z(t). Simple stats using one of the binning methods.

-- integrated raw t: will crash of depth inversions!
  Simply integrate t observations from surface to bin boundaries, stopping if
  a data gap of 10+m.

-- integrated bin t: will crash of depth inversions!
  Integrate t interpolated to bin boundaries, from surface to bin boundaries.
  Presently no gap criterion, so linear interolation could give poor
  results for gappy or non-filled low-res casts.

-- dtdz: apply binning method then calc dt/dz between bin values (so dtdz
  applies between bin centres and is allocated to the boundaries between those
  bins - so its a bit tricky to work out how to define the bins to end of with
  the desired result.

-- dzdt: same as for dtdz, except z gradient in t space.