Cropland PAtteRNS: Parallel Dimensional Attention Networks and Attention to Dataset Disparity for Crop Segmentation in Satellite Imagery Time Series Data
The paper says crop-segmentation benchmarks are being distorted by dataset design choices, not just model quality.
PAtteRNS splits self-attention across temporal, spectral, and spatial dimensions of Sentinel-2 time-series imagery, then recombines that with convolutional refinement. The authors report it beats compared crop segmentation models across several metrics, with particular gains on parcel-boundary quality measured by Boundary IoU. They also argue that flawed class groupings and different tile-size variants make some dataset comparisons unfair or incomparable. Code, models, and dataset preparation guides are listed as available with the paper. ArXiv · AI/CL/LG's note
PAtteRNS splits self-attention across temporal, spectral, and spatial dimensions of Sentinel-2 time-series imagery, then recombines that with convolutional refinement. The authors report it beats compared crop segmentation models across several metrics, with particular gains on parcel-boundary quality measured by Boundary IoU. They also argue that flawed class groupings and different tile-size variants make some dataset comparisons unfair or incomparable. Code, models, and dataset preparation guides are listed as available with the paper. ArXiv · AI/CL/LG's note
score 4