Skip to content

Neural Networks

Architecture Overview

CANYON-B (and so canyonb) uses a committee of neural networks for each predicted parameter. Each network in the committee:

  1. Contains 2-3 hidden layers
  2. Uses hyperbolic tangent activation functions
  3. Is trained on slightly different datasets
  4. Has different initializations

Network Committee Advantages

  • Improved prediction stability
  • Better generalization
  • Uncertainty estimation
  • Robustness to outliers

Training Data

The networks are trained on high-quality reference data from:

  1. GO-SHIP cruises
  2. Time series stations
  3. Research expeditions

Parameter-Specific Networks

2 Parameter-Specific Networks

  1. Carbonate System

    • AT
    • CT
    • pH
    • pCO2

  2. Nutrients

    • NO3
    • PO4
    • SiOH4

Input Preprocessing

Before neural network prediction

  1. Inputs are normalized
  2. Geographic coordinates are adjusted
  3. Time is converted to decimal years
  4. Arrays are reshaped as needed

Output Processing

After neural network prediction

  1. Results are denormalized
  2. Uncertainties are calculated
  3. Arrays are reshaped to match input
  4. Results are packaged into dictionary