Résumé
To face the ever-increasing threats on ecosystems and their biodiversity, protected areas are expanding worldwide but assessing their effects requires advanced monitoring methods. Environmental DNA (eDNA) metabarcoding can reveal local biodiversity from genetic material released in the environment, offering ecosystem-health insights without invasive or destructive surveys. Yet, to harness the full potential of eDNA data in ecosystem assessment, robust and direct links between raw eDNA sequences and protection status remain to be learned with machine learning. Here, we present the first fully end-to-end framework learning and predicting a continuous protection gradient directly from raw eDNA sequences, bypassing taxonomic assignment or precomputed biodiversity indices. Our framework combines contrastive self-supervised pre-training to extract meaningful embeddings from eDNA sequences with a neural-network classifier to predict the level of protection. Applied to eDNA fish surveys on the Mediterranean coast, we show that raw sequences encode protection status along a gradient from unprotected to fully protected areas. This general framework can facilitate hypothesis testing on disturbance gradients and indicate whether a site is degraded or effectively protected, providing a scalable tool for ecosystem monitoring and adaptive management.