Mapping fractional vegetation cover in UAS RGB and multispectral imagery in semi-arid Australian ecosystems using CNN-based semantic segmentation

Abstract Context Monitoring fractional vegetation cover (FVC) is crucial for assessing ecosystem health and sustainably managing semi-arid rangelands. Field-based methods are resource-intensive, while moderate-resolution satellite imagery lacks the spatial detail needed to capture ecosystem complexity. This study leverages centimetre-scale UAS RGB and multispectral imagery with convolutional neural networks (CNN)-based U-net segmentation framework to address these challenges. Objectives We tested the integration of UAS multispectral imagery and CNN models for mapping FVC in semi-arid rangelands with varying vegetation types. Methods We trained and evaluated site-specific and generic experimental U-net CNN models to classify FVC into five classes: bare ground (BE), photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV), shadow (SI), and water (WI). Data from three semi-arid sites with varying vegetation structures were used. Spatial block cross-validation was applied to reduce spatial autocorrelation, and performance was assessed using accuracy, F1-score, and IoU metrics. Results Site-specific models achieved high test performance, with Overall Accuracy (OA) scores of 89.56%, 88.11%, and 68.52% for the low, medium, and dense vegetation sites, respectively. The reduced accuracy in the dense site was attributed to spectral ambiguity, complex vegetation structure, and limited reference data. The generic model showed poor generalisation, reaching only 28.48% OA. Performance improved under modified conditions: Experiment A (excluding the water class) reached 82.65% OA, while Experiment B (excluding water and applying data augmentation) achieved 77.53%. Conclusion Site-specific CNN models mapped FVC components at centimetre-scale resolution, demonstrating robustness for specific vegetation types. This study establishes a workflow for high-resolution vegetation mapping and highlights the need for expanded training data, advanced augmentation, and complementary sensors (e.g., LiDAR, SAR, hyperspectral) to enhance generalisation and transferability for ecosystem monitoring.