Comparative Performance of Multi-Spectral Vegetation Indices for Phenology-Based Rapeseed Classification

Abstract

Rapeseed (Brassica napus L.) is a globally important oilseed crop, and its accurate monitoring through remote sensing is crucial for timely and informed agricultural decision-making. This study aimed to evaluate the efficiency of 21 vegetation indices (VIs), including both commonly used and flower-sensitive indices, for classifying rapeseed fields using time-series Sentinel-2 imagery. We utilized 50 Sentinel-2 images acquired throughout the growing season to capture phenological variation. A supervised classification approach based on the Random Forest algorithm was implemented to distinguish rapeseed from non-rapeseed pixels. The results revealed that VIs sensitive to changes in green and red reflectance (e.g., GRVI, VARI) and those contrasting green and blue reflectance (e.g., NDYI) performed best, achieving overall accuracy (OA) values up to 0.99, Kappa coefficients around 0.97, and F1 scores near 0.97. These top-performing indices also exhibited the lowest false positive and false negative rates. In contrast, traditional biomass-oriented indices such as CI and MSAVI performed poorly, with lower OA (~0.94) and significantly higher false positive rates, likely due to their insensitivity to the spectral effects of flowering. Our findings confirm that flower-sensitive indices are better suited for capturing the phenological signals of rapeseed flowering, especially those in the visible spectrum, while indices primarily relying on NIR and red-edge features are less effective under flowering conditions. We conclude that a phenology-based classification approach, when supported by well-selected training data and appropriate indices, can yield highly accurate results. We recommend that future studies adopt the most effective indices identified in this study—particularly GRVI, VARI and NDYI—for operational monitoring and mapping of rapeseed fields using Sentinel-2 data.