Fig. 5: Performance chart for a set of various data-driven techniques used to estimate eddy heat fluxes from sea surface height snapshots.
From: Deep learning to infer eddy heat fluxes from sea surface height patterns of mesoscale turbulence
The y-axis represents their obtained skill and each column has a denoted value of R2 that reflects the fraction of the total eddy flux variance explained by the fit. The following techniques are shown: Linear Regression, Random Forrest, Principle Component Analysis (PCA), Support Vector Machine (SVM), Fully Connected shallow Neural Network (FC NN), Convolutional Neural Networks with sequential propagation (VGGNet) and with residual connections (ResNet). The techniques are described in Methods. Convolutional neural networks (dark blue) significantly outperform other methods (cyan) as they are optimized to extract the most informative eddy patterns. Note that the FC NN and VGGNet training time is O(1 cpu hour), whereas ResNet takes O(1000 cpu hours) to train.
