Convolutional Neural Networks for the Segmentation of Oceanic Eddies from Altimetric Maps

(Technical report)

"Convolutional Neural Networks for the Segmentation of Oceanic Eddies from Altimetric Maps". Preprint can be found here

I already made public some jupyter notebooks and data to let anyone start using it.

EddyNet

(IGARSS conference paper)

EddyNet: A Deep Neural Network For Pixel-Wise Classification of Oceanic Eddies

This is the supplementary material of the publication "EddyNet: A Deep Neural Network For Pixel-Wise Classification of Oceanic Eddies", from R. Lguensat et al., accepted as an oral presentation for IGARSS2018. Pre-print at: https://arxiv.org/abs/1711.03954

Eddynet is an U-Net like architecture (a convolutional encoder-decoder followed by a pixel-wise classification layer + skip connections).

Paper main messages:

• A deep neural net that "emulates" the result of a geometry based and expert based method
• Comparing EddyNet with a version where we use SELU activation function (EddyNet_S). Replacing directly ReLU+BN with SELU resulted in a noisy loss and hurted the performance, we then kept BN after maxpooling, transposed deconvolution and concatenation.
• For this multiclass classification problem, we use (1-mean dice coefficient) as a loss function instead of the categorical cross entropy loss
• Eddynet is easily modulable and can be used for further studies such as adding new information (e.g. Sea Surface Temperature), or training with another ground truth.

Some examples of the segmentation

---

---

Note: Like all the Eddy detection algorithms (there is no consensus on the best method to use), the ground truth is not 100% perfect, very few eddies are missed. Still working on an improved dataset and I encourage other researchers to do so.