SkinToSSTNet

The file pix2pix.py implements a generative network that map satellite observation of subskin Sea Surface Temperature (SST) to the first vertical model level, and viceversa. A tutorial on this kind of network can be found here (last access: June 7, 2024). This network has been used in the paper Assimilation of Diurnal Satellite Retrieval of Sea Surface Temperature with Convolutional Neural Network, M. Broccoli and A. Cipollone 2025 Mach. Learn.: Earth 1 015007, DOI 10.1088/3049-4753/adfb7e; the interested reader is referred to the manuscript for further details.

How to use the network

The network receives five input features, which must be given in the following order:

1. (sub)skin SST
2. wind speed
3. mask
4. latitude grid
5. longitude grid

and it outputs only one field of temperature:

6. SST

As described in the paper, the datasets used for training are the followings:

1. subskin SST from GHRSST https://doi.org/10.5067/GHGMB-3CO02
2. wind speed from the same above product
3. mask with 1 where there is a satellite measurement, 0 for land and clouds
4. rasterized latitude at 5 degree resolution
5. rasterized longitude at 5 degree resolution
6. ESA SST CCI and C3S https://doi.org/10.48670/moi-00169

Preprocessing

The network input shape is (720, 1440, 5), so that the datasets must be downsampled to 0.25 degree resolution. Training was performed on diurnal retrievals only, so that from the satellite measurements only the one obtained during daytime are to be kept. Anomalies were computed over the training dataset, i.e. the mean states of the training set was subtracted to each input and output feature, and the data was then scaled to the range [0,1]. To use the network, it is essential to reproduce these steps. To this end, mean values of our training dataset can be found on 10.5281/zenodo.11520481 in the directory preprocessing_setup, together with min and max values. The data can then be scaled to [0,1] for instance with the following function

def scaling(data, min, max):
    return (data - min)/(max - min)

Network predictions

The network can be used to project the subskin SST to first level SST (forward network), or also first level SST to subskin SST (inverse network). For both cases, we provide pre-trained weights of the network on 10.5281/zenodo.11520481, in the directories forward_network_weights and inverse_network_weights respectively. After downloading the weights, they can be loaded as follows:

# load pre-trained weights
weights_path = "./forward_network_weights" # to load forward network weights
# or
# weights_path = "./inverse_network_weights" # to load inverse network weights
checkpoint_path = weights_path + "/cp_epoch_{epoch:04d}.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
latest = tf.train.latest_checkpoint(checkpoint_dir)
print(latest)
gan.load_weights(latest)

The predictions of the network are obtained by calling the generator on the input data.

Postprocessing

To convert the network predictions back to Celsius degree, use

def undo_scaling(data, min, max):
    return data*(max - min) + min

and then add the corresponding (subskin or SST) mean value to obtain the full field.