Crop Classification under Temporal Shift through Sentinel-1 / Sentinel-2 Fusion

From September 2025 to August 2026, I held a Temporary Teaching and Research Assistant position at the Université du Littoral Côte d'Opale, under the responsibility of Prof. Émilie Poisson Caillault and Prof. Gilles Roussel. My research was carried out at LISIC, on the Saint-Omer site.

This work extends my PhD research along two lines: the extension of the weighted recognition approach, exploring new, possibly nonlinear, weighting functions for the most discriminative spectral bands; and crop classification under temporal shift, presented here.

Accurate crop-type mapping from satellite imagery is an essential task for large-scale agricultural monitoring. Classification models, however, suffer from temporal shift: the spectral signatures of a given crop vary considerably from one year to the next due to environmental and phenological factors.

This work explores dual-branch convolutional neural networks for the multimodal fusion of multispectral (Sentinel-2) and synthetic aperture radar (Sentinel-1) patches, with an adaptive feature fusion mechanism, trained on the years 2020 to 2023 and evaluated on 2024.

• A spatial reweighting strategy based on spectral prototypes, applied as a preprocessing step to improve robustness to mixed spectra and noisy observations.
• With this reweighting, a 275,000-parameter network reaches 72.63% accuracy, compared with 55.0% without reweighting and 62.6% for a ResNet18 with more than 11.5 million parameters.
• Identification of an open issue: on an unseen year, the macro-F1 score remains at 0.5720, which motivates further work on domain adaptation and out-of-distribution detection.