This work contributes to the environmental monitoring and sustainable management of Mediterranean agroforestry ecosystems by assessing the vegetative and structural status of Portuguese Quercus suber montados through the integration of multispectral UAV surveys and deep learning models. In a context of increasing climate variability, rapid, accessible, and automatable methods for crown monitoring are essential for timely evaluations of cork oak ecosystems status and dynamics. The case study was conducted at Monte Fava (Portugal), a Q. suber genetic field trial suited for investigating relationships between structural and physiological crown traits and potential genetic, adaptive, and functional differences among provenances. In this context, an automated tool for crown segmentation and metric characterization enables rapid, repeatable, and non-destructive monitoring of tree growth and vigour, and also response to environmental variability.

Methodologically, the work integrates multispectral UAV data, digital photogrammetry, manual crown annotation, and deep learning modelling. Images acquired with a DJI Phantom 4 Multispectral RTK drone were processed in WebODM and georeferenced in EPSG:3763. Approximately 600 crowns were manually digitized and used as ground truth for training a lightweight convolutional neural network (U-Net-light), balancing accuracy, computational efficiency, and operational accessibility even under limited hardware resources.

The network was tested on multiple multiband and single band to evaluate spectral contribution and model robustness under different input configurations. Despite its simplified architecture, multiband models achieved high performance (IoU up to 0.82), demonstrating good generalization and temporal robustness. This confirms that lightweight, well-trained networks are a valid alternative to more complex models when computational cost, transferability, and ease of use are priorities. Inference on georeferenced orthophotos produced accurate, morphologically coherent crown masks from which structural parameters, such as crown area, were derived. Estimated crown surfaces, mainly ranging between 70 and 110 m², are consistent with the expected dimensions for adult Q. suber individuals. The compactness of the crowns and the regularity of the segmented shapes confirm model reliability for both spatial vegetation classification and metric characterization. This pipeline can accommodate main vegetation indices across the orthophoto, enabling integration of geometric crown information with multispectral indicators of vegetative vigour and photosynthetic function.

The combination of U-Net-light, multispectral UAV data, and morphometric analysis provides an original and scalable approach to automated Q. suber crown characterization, suitable for precision forest monitoring. Its application at Monte Fava genetic trial demonstrates its value for comparative assessment, phenotypic plasticity analysis, and climate-resilient montado management. Beyond accurate crown segmentation, the methodology establishes an operational bridge between spatial data, vegetation structure and vigour indicators, offering a powerful tool for precision forestry in high-value ecological systems.