Oak Dieback Detection: Sentinel-2 Time Series vs. Multimodal Geospatial Foundation Model Embeddings
Year: 2026
Type: article
Abstract: Oak dieback from recurrent drought manifests as a subtle, gradual alteration of canopy vigor that is difficult to detect. Expert-designed Sentinel-2 (S2) time-series indices — continuum-removal (CR) indices targeting shortwave-infrared water and red-edge chlorophyll absorption — are effective but require domain expertise and heavy preprocessing. Pre-trained geospatial foundation models (GFMs) promise general-purpose embeddings with minimal engineering, yet their ability to resolve within-class biophysical gradients remains unexplored. This study compares the TESSERA and AlphaEarth embeddings with an operational oak-dieback detection framework based on expert-designed S2 time-series indices in an oak-dominated study area in the northern half of France centred on the Centre-Val de Loire region. Using in situ observations collected from 2017 to 2022 on 2 030 oak plots, we evaluate how well these approaches classify forest health into three categories (healthy, declining, very declining). We further assess the influence of temporal resolution, of aggregation strategies (pixel-based, plot-level majority voting, and plot-mean feature aggregation), and of computational cost on dieback-mapping performance. The S2 time-series indices and TESSERA both outperform AlphaEarth in all configurations, with the S2 indices best in the pixel-based setting under the Random Forest classifier used as the primary comparison; this ranking is not representation-intrinsic, however, as it reverses under a logistic-regression linear probe, where both embeddings significantly outperform the S2 indices. The S2-indices advantage is also conditional on the survey years being represented in the training data: under a strict temporal hold-out, the intra-annual indices fail to transfer forward to unseen survey years, where the annual TESSERA embeddings degrade more gracefully. The hybrid configuration (expert-designed S2 indices combined with GFM embeddings) provides limited additional benefit, with AlphaEarth the more useful complement. The results highlight the importance of intra-annual temporal resolution and confirm the role of seasonal dynamics in detecting forest decline. Despite their lower accuracy, the GFM embeddings offer a substantial operational advantage in computational cost, delivering results in minutes rather than the hours required to process the S2 time-series indices.
Language: English
Subfield: Ecology
Field: Environmental Science
Domain: Physical Sciences
SDG: Life on land
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7220846005
