On the synergitic use of the L2 and L3 UAV lidar infrared intensity with visible and surface models for habitat classification: A case study on a complex socio-ecological littoral (Brittany, France)

Coastal socio-ecological interfaces—where beaches, dunes, salt marshes and mudflats meet human infrastructure—rank among the most biodiverse yet rapidly changing environments. Mapping their habitats is hampered by spectral confusion between radiometrically similar surfaces: wet versus dry sand, mud versus low-marsh vegetation, and the subtle zonation of upper and lower salt marsh. Lidar intensity provides complementary information to the blue–green–red (BGR) imagery from UAV surveys, yet the contribution of 905 nm (DJI Zenmuse L2) and 1550 nm (DJI Zenmuse L3) intensity to classification accuracy remains, to our knowledge, unquantified.

At Les Chevrets (Côte d'Émeraude, Brittany, France), a beach–dune–saltmarsh–mudflat continuum, ten habitat classes were targeted: wet sand, dry sand, mud, lawn, dune vegetation, high-saltmarsh vegetation, low-saltmarsh vegetation, road, roof and tree. On a 4000 × 3000-pixel scene at 0.1 m resolution, six input stacks were compared: BGR; BGR with 905 nm intensity; BGR with 1550 nm intensity; BGR with 905 nm and 1550 nm intensity; BGR with a digital surface model (DSM); and BGR+DSM with 905 nm and 1550 nm intensity. Each was classified with three per-pixel spectral classifiers—maximum likelihood, random forest and support-vector machine—and a contextual TensorFlow-based U-Net encoder–decoder performing pixel-wise semantic segmentation. Accuracy was assessed against independent reference samples using confusion matrices, overall accuracy and kappa.

Preliminary maximum-likelihood results improved overall accuracy over BGR alone (89.39%) by 4.23%, 4.44% and 6.31% with 905 nm, 1550 nm and combined intensity, respectively. Adding the DSM to BGR raised accuracy by 9.85%, while BGR+DSM+905 nm+1550 nm yielded a 10.28% gain.

The synergistic use of 905 nm and 1550 nm lidar intensity constitutes a valuable complementary band for fine-grained coastal habitat mapping; combined with surface-model and non-parametric or deep-learning classifiers, it substantially improves the delineation of spectrally ambiguous intertidal and supratidal habitats, supporting the monitoring and management of complex socio-ecological systems.