whigg/w-net

A Deep Network for Simultaneous Identification of Gulf Stream and Rings from Concurrent Satellite Images of Sea Surface Temperature and Height

W-NET

From the literature we found that Sea Surface Temperature (SST) and Sea Surface Height (SSH) are the primary oceanographic variables that characterize the gulf stream and rings. We designed a network(W-Net) that could use SST and SSH data for automated and simultaneous identification of these synoptic ocean features.

Main Libraries required

1. pytorch
2. Matplotlib
3. OpenCv
4. Pillow

Download SST and SSH data

Our dataset consists of Sea Surface Temperature (SST) maps, Sea Surface Height (SSH) maps, and the manual annotations of gulf stream and eddies by an expert. Our focus area is the region bounded by 85◦W to 55◦W and 20◦N to 45◦N. You can download the SST and SSH data using the steps described in the next section, but the maunal annotations of gulf stream and eddies are not publicaly available so we will skip that here.

SST data

We used the Level 4 Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) product as our SST input. This data is available at 5.5km resolution. We also tried 1km SST data but did not get any improvement in the network's performance.

link for OSTIA data : https://podaac-tools.jpl.nasa.gov/drive/files/allData/ghrsst/data/L4/GLOB/UKMO/OSTIA/ link for 1km SST data : https://podaac tools.jpl.nasa.gov/drive/files/allData/ghrsst/data/L4/GLOB/JPL_OUROCEAN/G1SST/

1. Mount the podaac drive on your local machine using: sudo mount.davfs web_link /your_directory
2. Run Download_data/SST_Poodac.py to download the Net-CDF files
3. Run Download_data/SST_nc_to_colormap.py to convert Net-CDF files to images

SSH data

We used the gridded product of sea-level anomaly provided by copernicus marine environmental monitoring service (CMEMS) as our SSH input. This data is available at 27km resolution. One needs to manually download the SSH data from the link below-

1. Download all the NC files from year 2014 to 2018 from the below link- https://cds.climate.copernicus.eu/cdsapp#!/dataset/satellite-sea-level-global?tab=overview
2. Run Download_data/SSH_nc_to_colormap.py to convert Net-CDF files to images

Training

To train the model run train.py
train.py uses load_data.py to load the data.
loss.py file contains differnt loss function used for training. One can add custom loss functions as well and use these new losses to train the model.
train.py calls w_net_model_r_1.py as it contains the main W-Net architecture.

Here are a few thing that user can specify in the train.py to perform several experiments-

1. Number of classes and labels of classes. The network can be trained on all four labels (Warm eddies, Cold eddies, Gulf Stream and background) or it can be trained on a subset of these labels.
2. Loss function (crossentropy, Dice or a mix of both)
3. Type of model (W-Net, Res-W-Net, Y-Net, U-Net-SST, U-Net-SSH)
4. Data (One can use full data, winter months data or summer months data)
   The train.py saves the model in w-net.pth, which can be used for testing and further evaluation.

Test

To test the model run Test/test.py
test.py gives the final test accuracies for all the labels. It also saves the images with overlapping ground-truth and perdicted segments for a better visual inference.

Evaluation metrics

Commonly image segmentation is done on natural images, for which IOU and dice coefficient are used for both evaluation and training. However, in our application, we are interested in the dynamics of gulf stream and rings. Thus, different dynamic-inspired metrics are designed to evaluate the performance of the deep network.

Eddy evaluation

We devise metrics that compare the size, centroid, and count between the ground truth and the network prediction for eddies. All the eddy evaluation metrics are calculated in detect_eddies.py

Gulf stream evaluation

One key challenge in automated detection of the Gulf Stream is to capture its meandering path accurately. We first converted the gulf stream segments to its centerline using a morphological thinning operation using thinning.py. Then we computed path length difference and several curve difference metrics.

Codes for all the Gulfstream evaluation metrics are saved in the folder 'Gulf stream evaluation metrics'.
Gulf stream evaluation metrics/Hausdorff_distance.py computes the Hausdorff distance, Mean curve distance and Median curve distance.
Gulf stream evaluation metrics/path_length.py computs the mean and median path length difference between the ground truth and the predicted gulf stream centerline.

Reference

[1] D. Lambhate, R. Sharma, J. Clark, A. Gangopadhyay and D. Subramani, "W-Net: A Deep Network for Simultaneous Identification of Gulf Stream and Rings From Concurrent Satellite Images of Sea Surface Temperature and Height," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2021.3096202.