Detection and monitoring of Melampsora spp. Damage in multiclonal poplar plantations coupling biophysical models and Sentinel-2 time series

Abstract

Climate change is dramatically shifting the distribution and prevalence of pests and diseases, posing significant threats to global forest ecosystems. Poplar plantations, particularly multiclonal ones, are highly vulnerable to pathogen-driven diseases such as leaf rust caused by Melampsora spp. In this study, we developed three machine learning (ML) detection models (DMs) for identifying rust-affected poplar trees coupling Sentinel-2 time series and the PROSAIL radiative transfer model. For each DM, three ML algorithms (support vector machines, random forests, and neural networks) were trained using in situ leaf rust inspections as reference data, and the following inputs: (i) inverted plant traits retrieved from the PROSAIL model, (ii) key spectral indices derived from Sentinel-2 time series, and (iii) a combination of both plant traits and indices from Sentinel-2 images. The best-performing DM, which combined plant traits and spectral indices, achieved an overall accuracy of 89.5 % (Kappa = 0.78) across three tested ML algorithms. Relative importance analysis highlighted chlorophylls (21 %), carotenoids (16 %), and leaf water content (11 %) as the most critical variables for rust detection. This study shows the potential of combining biophysical models with Sentinel-2 imagery for precise and scalable rust detection in multiclonal poplar plantations. Our approach also highlights how key plant traits, such as chlorophyll, carotenoids, and leaf water content, vary across poplar clones, offering valuable insights for forest management and conservation strategies in the context of climate change. The framework we propose is adaptable and transferable to different regions and conditions, enhancing disease monitoring and forest health management. Its robustness is further supported by external validation using the ANGERS spectral database, confirming the physiological relevance of the retrieved traits.