Repository for Unsupervised Inpainting for Occluded Sea Surface Temperature Sequences (Yin et al., 2019) presented at The 9th International Workshop on Climate Informatics in October 2019.
Sea Surface Temperature (SST) data remotely detected from satellites are inevitably occluded by clouds or rains. To learn the prior of SST data and complete the missing part, inpainting approaches in machine learning often require ground truth in the occluded area, which is not applicable for remote SST. In this paper, we propose an inpainting approach based on generative adversarial networks (GANs) for occluded SST. It is capable to complete the occluded area without any supervision on ground truth data of the occluded area. Our framework is evaluated on real SST data with simulated cloud masks.
To download the cloud dataset, please check download link.
To download the SST dataset, please sign in to Copernicus Marine environment monitoring service, check the product, choose GLOBAL-ANALYSIS-FORECAST-PHY-001-024-HOURLY-T-U-V-SSH, and select corresponding regions. For training and validation sets, select the period from 2018-01-01 00:30 to 2018-12-31 23:30 between latitude 20°~25.25°N (+20°~+25.25°) and longitude 34.75°~40°W (-40°~-34.75°). For test set, select the period from 2019-01-01 00:30 to 2019-06-30 23:30 between latitude 14.75°~20°S (-20°~-14.75°) and longitude 14.75°~20°W (-20°~-14.75°).
This code requires Python 3, and we recommend a version higher than 3.6.
To install all requirements, run
pip3 install -r requirements.txt
To run experiment, set cloud dataset path and SST dataset paths in src/unisst.yaml and experiment storage path in run.py, then run
bash train.sh unisst 0
on GPU cuda:0, you can use multiple GPU for larger batch sizes, for example
bash train.sh unisst 0,1,2,3
We shreshold Liquid Water Path (LWP, in unit of g/m2) to generate clouds at different occlusion rates. Here are some samples from test sets. (Each sample from top to bottom: unknown ground truth, observation, recovered sequence)
| LWP Threshold | Sample | LWP Threshold | Sample |
|---|---|---|---|
| 55 g/m2 |   |
70 g/m2 |   |
| 60 g/m2 |   |
75 g/m2 |   |
| 65 g/m2 |   |
80 g/m2 |   |