Reproduce “U-TILISE: A Sequence-to-sequence Model for Cloud Removal in Optical Satellite Time Series” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Corinne Stucker, Vivien Sainte Fare Garnot, Konrad Schindler
U-TILISE reconstructs complete multispectral optical satellite time series by combining a convolutional U-Net encoder-decoder with lightweight temporal self-attention. On EarthNet2021, it improves over linear interpolation on masked pixels, reaching 0.0086 MAE, 38.29 dB PSNR, and 0.970 SSIM in-distribution, while retaining the original temporal sequence rather than producing a single cloud-free composite.
Key insights
- Explicit temporal modeling is central: removing the temporal encoder raises ablation MAE to 0.0339 and lowers PSNR to 26.49 dB, compared with 0.0100 MAE and 36.82 dB for the corresponding full configuration.
- Calendar-based or relative temporal encoding helps the model account for irregular sampling and seasonal patterns; attention-weighted skip connections further preserve spatial detail.
- Optical temporal context alone can provide useful reconstructions, while adding SAR in an auxiliary experiment yields only marginal gains, increasing PSNR from 33.77 to 33.94 dB.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.