Permutation‑Invariant PIUnet Boosts Multi‑Temporal Satellite Image Super‑Resolution

Problem

Satellite sensors provide limited spatial resolution, which hampers precision agriculture, disaster monitoring, and sustainable‑development applications. Multi‑temporal low‑resolution observations can increase detail, but existing deep‑learning methods assume that the order of the input images matters.

Temporal permutation invariance

Temporal variations in satellite imagery arise from clouds, illumination changes, seasonal cycles, and human activity, making the sequence unpredictable. Unlike video, where frame order encodes motion, the ordering of multi‑temporal satellite images does not carry useful information for super‑resolution. Therefore a model that produces identical outputs for any permutation of the input set is desirable.

PIUnet architecture

PIUnet (Permutation‑Invariant and Uncertainty Network) enforces permutation invariance through equivariant layers followed by average‑pooling over the time dimension.

TEFA (Temporal‑Equivariant Feature Attention) shares convolutional kernels across time, extracts spatial features, and mixes them with a self‑attention mechanism that is designed to be permutation‑equivariant (it computes all pairwise correlations without positional encodings).

TERN (Temporal‑Equivariant Registration Network) is an equivariant version of the RegNet block used in DeepSUM. It computes adaptive filters as functions of the input and registers all images simultaneously, avoiding the need for a reference frame.

After processing each temporal instant with TEFA and TERN, an average‑pooling operation collapses the time axis, guaranteeing identical outputs for any input ordering.

Experimental results

Evaluation on the Proba‑V dataset shows:

State‑of‑the‑art super‑resolution quality comparable to DeepSUM while using only 25 % of the training data.

Significantly lower inference time than ensemble‑based approaches (PIUnet runs much faster, as illustrated by the timing comparison figures).

Higher computational efficiency because the model does not waste capacity learning to handle permutations.

AI4EO agricultural challenge

A variant of PIUnet equipped with a segmentation head processed multiple 10 m Sentinel‑2 images and produced 2.5 m super‑resolved segmentation maps, satisfying the competition requirement and winning the challenge.

Conclusion

Permutation invariance is a fundamental property for multi‑temporal satellite image super‑resolution. By building equivariant modules (TEFA, TERN) and aggregating over time, PIUnet demonstrates that models can achieve higher accuracy with less data and faster inference.

References

PIUnet paper: https://arxiv.org/abs/2105.12409

PIUnet code: https://github.com/diegovalsesia/piunet