Reproduce “Panoptic Segmentation of Satellite Image Time Series with Convolutional Temporal Attention Networks” 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
Vivien Sainte Fare Garnot, Loic Landrieu
The paper introduces a panoptic segmentation framework for satellite image time series (SITS) that identifies crop types and delineates individual agricultural parcels. Researchers developed U-TAE, a spatio-temporal encoder with multi-scale temporal attention, and PaPs, a single-stage panoptic segmentation head, achieving a 43.8 Panoptic Quality on the newly released PASTIS dataset.
Key insights
- Multi-scale temporal attention, where low-resolution attention masks are interpolated to higher resolutions and applied with channel grouping, is crucial for effectively extracting spatially adaptive and parcel-specific temporal features from SITS.
- Group Normalization is essential for SITS processing, outperforming Batch Normalization by a significant margin due to the non-identically distributed temporal samples and varying acquisition conditions.
- The temporal dimension in SITS is indispensable for accurate crop classification and parcel delineation, with models trained on single acquisitions showing drastically reduced performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.