A CNN with spectral-channel adaptation for interpretable and edge-ready fractional vegetation cover mapping using UAS multispectral and imaging spectroscopy, extended to 3D LiDAR structure
Year: 2026
Type: article
Abstract: An Earth Observation Adventure: From 2D Cover to 3D Structure This presentation, delivered at the Australian Earth Observation Forum 2026, explores automated fractional vegetation cover mapping using UAS multispectral (MicaSense) and imaging spectroscopy/hyperspectral (Specim FX10–FX17) imagery collected at TERN AusPlots sites on Calperum Station, South Australia. The research investigates a U-Net architecture incorporating spectral-channel adaptation, channel attention, and an NPV-weighted loss, evaluated using spatial block cross-validation. Within-site results achieved NPV F1-scores of 0.73 for multispectral and 0.46 for hyperspectral imagery, outperforming Random Forest, SS3FCN, and baseline U-Net models. NPV-weighted loss contributed the largest improvements. However, zero-shot transfer to an independent site revealed substantial domain-shift limitations, with NPV F1-scores of 0.12 or lower across all models. The presentation also highlights edge AI feasibility using TensorRT on an NVIDIA Jetson AGX Orin (1.70× inference speed-up with 99.96% prediction agreement; benchmarked but not yet flown) and ongoing research transferring 2D vegetation growth-form labels to LiDAR point clouds for 3D semantic segmentation. Together, these studies explore the opportunities and challenges of integrating spectral, spatial, and structural information for scalable ecosystem and biodiversity monitoring. Code repositories https://github.com/LNSOTOM/fvc_composition https://github.com/LNSOTOM/fvc_function https://github.com/LNSOTOM/fvc_structure Related publication: https://doi.org/10.1007/s10980-025-02193-y Keywords: Fractional vegetation cover; Non-photosynthetic vegetation; UAS; Multispectral; Hyperspectral; Deep learning; Domain shift; LiDAR; Edge AI; Drylands.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
SDG: Life on land
Open Access status: green
OpenAlex ID: W7236810873
