Multispectral imagery for forest conditions monitoring of a pilot watershed in the Vosges, France

Details

Supervisors

Defourny, Pierre ; Thomas, Maxime ; Kenda, Tom

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

Forest ecosystems are increasingly affected by climate-induced stressors, resulting in progressive forest dieback and declining ecosystem resilience. In this context, efficient large scale monitoring approaches are essential for supporting sustainable forest management and climate adaptation strategies. This study evaluated the potential of Sentinel-2 multispectral time-series imagery combined with machine learning techniques for detecting and monitoring forest dieback in the Doller Valley, France, while also identifying the most informative Sentinel-2 spectral properties associated with dieback dynamics. Sentinel-2 imagery acquired between 2016 and 2025 was integrated with field observations collected from 189 forest plots. A plot-level Weighted Dieback (WDB) score was derived from field measurements to represent canopy decline severity. Vegetation indices; NDVI, NDWI, IRECI, CR_RE, and CR_SWIR were calculated, from which spatial and temporal variability metrics were extracted. These features, together with WDB scores, were incorporated into a Random Forest (RF) classification framework for healthy, moderate, and severe dieback classes. The final RF model achieved an overall accuracy of 70%, with F1-scores of 0.75, 0.61, and 0.75 for healthy, moderate, and severe classes, respectively. The results demonstrated that the shortwave infrared (SWIR), red-edge, and red spectral regions were the most informative for forest dieback detection, with CR_SWIR, CR_RE, and IRECI showing the strongest sensitivity to canopy chlorophyll degradation and vegetation water stress. In contrast, NDVI alone was found insufficient for reliable dieback detection, as understory vegetation could compensate for canopy loss and mask actual forest decline. Statistical analyses revealed clear spectral separability between healthy and severely affected plots, whereas moderate dieback conditions exhibited substantial spectral overlap with healthy forests. However, the inclusion of temporal variability and spatial heterogeneity metrics improved the detection of moderate dieback conditions, highlighting the importance of time-series information for identifying gradual canopy decline processes. Annual dieback prediction maps further revealed pronounced spatial and temporal variability between 2018 and 2025, with the highest severe dieback extent observed in 2018, followed by a gradual reduction toward 2024. Overall, the findings demonstrate that Sentinel-2 imagery combined with machine learning provides an effective and scalable framework for forest dieback monitoring in forest ecosystems. The developed approach can be further improved through the integration of higher-resolution datasets and application across different forest types and ecological regions, contributing toward more generalized and transferable remote sensing based forest health monitoring frameworks.