Principle

  1. Locality is important: make use of local region.
  2. Data, especially observed SST data, is imprecise due to the range of processing steps carried out. (And resolution scale effects are superimposed on this.)
  3. Nonlinear mapping - here, multilayer perceptron - caters for complex decision surfaces.
  4. To validate results, we take the data as consisting of: (i) learning set; (ii) validation set; and (iii) missing values, generalization set.
  5. We use as a preprocessing phase the redundant B3 spline à trous wavelet transform. Hence: wavelet nonlinear regression.
  6. Neighborhood information is accounted for (i) in inputs to multilayer perceptron mapping, and in (ii) resolution scale, which is based on low-pass filtering.
  7. Currents and heating may exist on different scales: how important is a multiscale analysis?
  8. In fact, we look at how an OGCM model reconstruction can be checked against the observed data at a range of different resolution scales.