Multispectral and hyperspectral remote sensing for winter cover crop identification, mixture and species characterisation in Wallonia, Belgium

Details

Supervisors

Defourny, Pierre ; Descarpentries, Louis

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

enWinter cover crops (WCCs) deliver key agro-environmental benefits that depend strongly on the composition of the species and mixtures used. Despite growing policy interest under the Common Agricultural Policy (CAP), spatially explicit information on WCC composition at a regional scale remains scarce. This thesis evaluates the potential of Sentinel-2 (S2) multispectral (MSI) and EnMAP hyperspectral (HSI) imagery for monitoring WCC in the Hesbaye region of Wallonia, Belgium, across three hierarchical classification tasks: WCC presence versus absence; pure versus mixed WCC stands; and dominant species composition. Using a reference dataset encompassing the agronomic diversity of WCCs in the region — pure stands, binary mixtures, and complex multi-species combinations dominated by mustard, phacelia, and companion species — three experimental setups were implemented within a consistent Random Forest (RF) framework: single-date EnMAP classification; a harmonised single-date EnMAP–S2 comparison; and S2 monthly composites spanning the WCC growing season (August 2025–March 2026). WCCs presence was mapped with high accuracy across all setups (OA: 80.72-86.04%, F1macro: 78.6-85.61%), confirming the viability of both sensors for operational soil cover monitoring. EnMAP outperformed S2 significantly for RF-1, suggesting that radiometric quality is as important as spectral sampling density. Distinguishing pure from mixed WCC stands remained moderately achievable (OA: 67.8-72.76%, F1macro: 62.81-65.55%), with consistently higher F1 scores for mixed than for pure classes across all setups. Species composition mapping was the most challenging task (OA: 39.48-44.54%, F1macro: 18.58-29.09%). Single-date EnMAP achieved moderate F1 scores only for phacelia-dominated and pure mustard classes. No systematic sensor advantage over S2 was observed at this level as classification errors were dominated by class imbalance and within-class spectral variability rather than spectral resolution. These results demonstrate that S2 time series can reliably support the monitoring of broad WCC, while EnMAP-like HSI adds measurable value for WCC detection and spectral band diagnostics, but not yet for reliable species-level mapping under current data and acquisition constraints.