Cropland PAtteRNS: Parallel Dimensional Attention Networks and Attention to Dataset Disparity for Crop Segmentation in Satellite Imagery Time Series Data

TL;DR

The paper introduces PAtteRNS, a hybrid transformer-convolutional model that applies self-attention independently across temporal, spectral, and spatial dimensions of Sentinel-2 satellite data using a parallel architecture to reduce computational cost. It demonstrates superior performance in crop segmentation, particularly in parcel boundary accuracy, and identifies critical issues in dataset design such as inconsistent class groupings and tile-size effects.

Why It Matters

Accurate crop segmentation is essential for agricultural monitoring and policy-making, and this work improves model efficiency and reliability while highlighting flaws in current dataset practices that undermine fair model evaluation and reproducibility.

Key Findings

• The parallel transformer architecture enables factorised attention across temporal, spectral, and spatial dimensions with lower computational overhead than standard triple-attention mechanisms.
• PAtteRNS achieves state-of-the-art results on PASTIS and MTLCC datasets, especially in boundary precision as measured by Boundary IoU.
• Dataset inconsistencies—such as non-standard class groupings and varying tile sizes—lead to non-comparable model performance, calling for standardized dataset construction practices.