Retrieving mangrove biochemical traits under oil pollution using hyperspectral data and transfer learning

Mangrove forests are essential in improving coastal resilience, biodiversity, and carbon sequestration. However, they are increasingly threatened by anthropogenic pressures, particularly pollution such as oil contamination, which can induce long-term physiological stress in mangrove vegetation. Yet, monitoring the health of such ecosystems across large areas is difficult due to the scarcity of ground-based data and the spatial resolution–coverage trade-off. In light of this, we propose a scalable deep learning framework that incorporates hyperspectral imagery and transfer learning to classify mangrove species and retrieve canopy-level biochemical traits, including pigments, sugars, and phenols. The framework is a Multi-Task Convolutional Neural Network (MTS-CNN) organized in a teacher–student architecture with pseudo-labeling. The teacher model, trained on airborne hyperspectral imagery at 0.5 m resolution, generates pseudo-labels and transfers its learned weights to student models developed on satellite datasets from PRISMA (5 m, pansharpened) and EnMAP (30 m), enabling knowledge transfer from high- to lower-resolution sensors. Despite differences in spatial resolution, the predicted biochemical patterns remained spatially consistent across sensors. Results show accuracies above 95% for airborne data, approximately 70% for PRISMA, and 60% for EnMAP. Spectral analysis additionally confirmed that the predicted classes retained key absorption features characteristic of the target biochemical compounds. Furthermore, the resulting maps could distinguish biochemical signatures associated with oil contamination from those caused by other stressors. These results demonstrate that cross-scale transfer learning can support broad monitoring of mangrove ecosystems by propagating biochemical information from high-resolution observations to satellite hyperspectral imagery.