Deep Learning for Invasive Species Mapping: Evaluating Semantic Segmentation of Spartina alterniflora Using Drone Imagery

The proliferation of invasive smooth cordgrass (Spartina alterniflora) within the Bay of Brest poses a severe ecological threat to coastal wetlands, disrupting sedimentary dynamics and endangering high-value native taxa such as Limonium humile. Within the Armorique Regional Natural Park (PNRA), ecosystem preservation has traditionally relied on manual photo-interpretation a labor-intensive approach limited by estuarine morphological complexity and the critical need for early invasion detection. This study evaluates the efficacy of coupling high-resolution Unmanned Aerial Vehicle (UAV) imagery with advanced semantic segmentation models to automate fine-scale habitat mapping across spectrally heterogeneous coastal environments.

Utilizing aerial data acquired in 2025 across the Daoulas estuary via EbeeX and Mavic 3E UAV platforms, integrated with precision GNSS-RTK ground surveys, we constructed a sub-metric reference dataset. This dataset benchmarks state-of-the-art deep learning architectures, specifically investigating the capacity of vision transformers (such as SegFormer and Swin-Unet) to accurately differentiate Spartina from native halophytic vegetation with high botanical fidelity.

The resulting framework establishes a robust, reproducible workflow for spatio-temporal monitoring that overcomes conventional fieldwork constraints. To foster open science and direct technology transfer, all developed models, workflows, and code have been released as open-source computational mapping solutions, notably via the SemanticSeg4EO framework. This open-access pipeline provides coastal stakeholders and PNRA managers with an accessible software environment to automate high-priority monitoring, thereby optimizing ecological restoration policies and guiding targeted, real-time invasive species eradication strategies.