A Modelling and Evaluation Framework for EuroCrops-Driven Sentinel-2 Crop Segmentation

arXiv:2606.00676v1

Announce Type: new

Abstract:

This study introduces a flexible pipeline designed to produce agricultural datasets suitable for semantic segmentation by integrating Sentinel-2 imagery with EuroCrops parcel-level annotations. The process converts diverse vector crop data into aligned multispectral image-mask pairs through a series of steps: label harmonization, selection of appropriate Sentinel-2 products, spatial alignment, rasterization, patch extraction, quality filtering, and class-aware sample selection. The resulting dataset comprises 67,337 patches sourced from five European nations, utilizing a simplified taxonomy of ten crop categories alongside a background class.

For model training, a four-level U-Net architecture equipped with Group Normalization was employed. The network was trained on ten Sentinel-2 spectral bands using a composite loss function that merges class-weighted cross-entropy with Dice loss. Internal evaluation on the EuroCrops-based test split yielded a mean Intersection over Union (mIoU) of 0.7665, pixel accuracy of 0.8693, and a mean class accuracy of 0.9072. These results underscore the significance of learned multi-scale spatial representations for crop segmentation, outperforming spectral and spatial-context Random Forest baselines.

External validation was conducted using unseen Belgian EuroCrops subsets, as well as the DACIA5 and PASTIS benchmarks. The findings reveal a notable performance disparity in external and cross-dataset scenarios, particularly when benchmarks differ in taxonomy, annotation protocols, spatial coverage, or temporal structure. While the model demonstrates reliable transferability to dominant, taxonomically aligned classes like maize and wheat, its effectiveness is constrained for certain minority classes and in the adapted single-date configuration of PASTIS. These outcomes illustrate both the capabilities and the constraints of leveraging EuroCrops-derived supervision for Sentinel-2 crop segmentation in the presence of realistic domain shifts.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC