Work

Out-of-year parcel identification using adversarial domain adaptation and multi-temporal satellite imagery to enhance vineyard health monitoring

Year: 2026

Type: article

Abstract: Accurate monitoring of vineyard parcel status is essential for precision viticulture and phytosanitary risk management, particularly for detecting abandoned parcels that may act as disease reservoirs for pathogens such as Flavescence dorée. Conventional monitoring methods are limited by restricted spatial coverage and low temporal frequency, making satellite remote sensing a key alternative for large-scale vineyard monitoring. This study addresses the challenging task of out-of-year vineyard parcel classification, which requires models to generalize across years with varying environmental and climatic conditions. We propose a deep learning framework based on conditional adversarial learning to address temporal domain adaptation in multi-year satellite data. The approach relies on freely available Sentinel-1 (radar) and Sentinel-2 (optical) imagery. The framework is evaluated through a real-world case study covering the Graves and Sauternes vineyards in western France. The dataset includes ground truth observations spanning five consecutive years (2020–2024), provided by the GDON du Sauternais et des Graves. Results show that the proposed method outperforms standard approaches in handling temporal domain shifts, with weighted F1-score improvements exceeding 5 percentage points overall and more than 10 percentage points for minority class detection, despite class imbalance and interannual environmental variability.

Source: International Journal of Digital Earth

DOI: 10.1080/17538947.2026.2634468

Authors: Anna Roussel, Samuel Alleaume, Adrien Billotte, Frédéric Fabre, Dino Ienco