Deep Learning and Remote Sensing for Salt-marsh Biomass Estimation in the Venice Lagoon

Salt marshes are vital coastal ecosystems that contribute to carbon sequestration, coastal protection, and biodiversity support. In the Venice Lagoon, these habitats face increasing pressure from sea-level rise, land subsidence, and human activities, which threaten their resilience and ecological function. Accurate and spatially distributed estimation of above ground biomass (AGB) is essential for monitoring salt-marsh health. This thesis presents a method to estimate AGB in the San Felice salt marsh using high-resolution multispectral data collected by unmanned aerial vehicles (UAVs) combined with artificial neural network (ANN) modelling. Multispectral data with 3 cm spatial resolution was acquired over the study area and processed to generate Orthomosaics and digital elevation models (DEMs). Field campaigns provided ground-truth biomass samples that were used to train and validate the ANN models. The study tested four network architectures and data augmentation strategies, including the in#tegration of synthetic samples to improve model robustness. Results indicate that a two-hidden-layer ANN incorporating spectral bands and elevation data yields the most accurate biomass predictions, with an R² of 0.89 and an RMSE of 135 g/m². The biomass spatial distribution predicted by the model corresponds well with known vegetation patterns, particularly along tidal channels. This research demonstrates the effectiveness of combining UAV multispectral data with deep learning to assess salt marsh biomass at fine spatial scales, addressing limitations of traditional methods based on coarser satellite data. The approach provides a scalable framework for detailed ecological monitoring in complex coastal environments. Further work is recommended to expand ground-truth datasets, integrate additional environmental variables, and validate the methodology across seasons and different marsh ecosystems.