Cropland PAtteRNS: Parallel Dimensional Attention Networks and Attention to Dataset Disparity for Crop Segmentation in Satellite Imagery Time Series Data
PAtteRNS factorizes temporal, spectral, and spatial attention for Sentinel-2 cropland segmentation.
Cropland PAtteRNS is a hybrid transformer-convolutional model for crop segmentation in Sentinel-2 multispectral satellite time series. It applies self-attention separately to temporal, spectral, and spatial dimensions and introduces a parallel transformer to reduce the cost of fully factorized attention. Ablations and comparisons on multiple tile-size variants of PASTIS and MTLCC show stronger crop-class segmentation, including parcel delineation measured by Boundary IoU. The authors also report that flawed class groupings and differing tile sizes can make dataset and model comparisons unreliable.
- Applies separate attention to time, spectrum, and space
- Uses a parallel transformer to cut triple-factorized attention cost
- Outperforms compared crop-segmentation models on PASTIS and MTLCC
- Finds class groupings and tile size distort fair comparisons
- Products
- Sentinel-2
- AI models
- PAtteRNS
From the original article
1 min read at arxiv.orgThe landscape of satellite imagery time series datasets and boundary-pushing architectures for cropland segmentation has never been richer. However, in this gold rush, important truths are being missed on both fronts, as a drive for the most novel concepts or the largest datasets pushes finer details to the side.
By Joseph Metcalfe · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.