A Multi-Mode Convolutional Neural Network (CNNMM) to reconstruct satellite-derived Chlorophyll-a time series in the global ocean from physical drivers
This repository contains the code of the model presented in the paper [A Multi-Mode Convolutional Neural Network to reconstruct Chlorophyll-a time series in the global ocean from physical drivers](Frontiers in Marine Science, 2023).
This repository contains the following PyTorch code:
- Implementation of the multi-mode CNNMM8 Chl time series regression from oceanic and atmospheric predictors :
Our model achieves the following performance on INDIGO Benchmark dataset :
| Model name | r2 | RMSE | Slope | Seas | Inter | N param | Time computation | Km travelled by car |
|---|---|---|---|---|---|---|---|---|
| CNNMM8 | 0.87 | 0.28 | 0.90 | 1.00 | 0.96 | 803 920 | 39 h | 8.9 |
See the paper for more details.
- torch==1.4.0
- torchvision==0.5.0
- numpy==1.18.1
- carbontracker==1.1.5
The benchmark formated dataset is available for download here. You can also find the required files to run the code in this repository. This dataset was built from the following source of data :
| Proxy used as predictors | Acronyme | Products | Initial spatio-temporal resolutions |
|---|---|---|---|
| Sea Surface Temperature | SST | Reyn_SmithOIv2 SST dataset | Monthly on a 1◦ × 1◦ spatial grid |
| Sea Level Anomaly | SLA | Ssalto/Duacs merged product of CNES/SALP project | Weekly on a 1/3◦ × 1/3◦ spatial grid |
| Zonal and Meridional surface winds | Uera, Vera | Atmospheric model reanalysis ERA interim 4 | Every 5-days on a 0.25◦ × 0.25◦ spatial grid |
| Zonal and Meridional surface total currents | u,v | OSCAR unfiltered satellite product | Every 5-days on a 0.25◦ × 0.25◦ spatial grid |
| Short-wave radiations | SW | NCEP/NCAR Numerical reanalysis | Daily on a 2◦ grid |
| Binary continental mask | mask | ||
| Bathymetry | bathy | GEBCO | 15 arc seconds |
Chl recontructed data over [2012-2015] :
animation_Chl_reconstructed_2012_2015.mov
Chl reference satellite data over [2012-2015] :
animation_Chl_satellite_reference_2012_2015.mov
Roussillon Joana, Fablet Ronan, Gorgues Thomas, Drumetz Lucas, Littaye Jean, Martinez Elodie (2023). A Multi-Mode Convolutional Neural Network to reconstruct satellite-derived Chlorophyll-a time series in the global ocean from physical drivers. Frontiers in marine science. doi: 10.3389/fmars.2023.1077623
Roussillon Joana, Fablet Ronan, Gorgues Thomas, Drumetz Lucas, Littaye Jean, Martinez Elodie (2022). satellIte phytoplaNkton Drivers In the Global Ocean over 1998-2015 (INDIGO Benchmark dataset). SEANOE. https://doi.org/10.17882/91910
