Fecha: 31/03/2025

Calificación: Sobresaliente Cum Laude

Disponible: https://riunet.upv.es/handle/10251/220802

Forests are critical global ecosystems, covering approximately 31% of land area and serving as key carbon sinks, biodiversity reservoirs, and climate regulators. They hold about 45% of terrestrial carbon and support around 80% of Earth’s above-ground biomass (Reichstein and Carvalhais, 2019). In recent decades, forests have faced increasing threats from wildfires. Although these natural processes are essential for nutrient cycling and habitat renewal, can cause severe impacts when intensified by human activity and climate change. This vulnerability is projected to increase under future climate scenarios, reinforcing the need for adaptive strategies that enhance forest resilience, carbon storage, and biodiversity (Lionello and Scarascia, 2018). In Mediterranean ecosystems, fire regimes are shaped by hot and dry summers and have been increasingly altered by land-use change, rural abandonment, and other anthropogenic pressures. In this context, effective forest management plays a critical role in mitigating wildfire risk by reducing fuel loads, maintaining landscape heterogeneity, or promoting structural conditions that limit fire spread and intensity.

Fire simulation models have become essential tools, helping to design fuel management strategies, plan protection infrastructure or guide land-use decisions (Mell et al., 2006). These models rely on the fire behaviour triangle (fuel, topography, and weather) and vary in complexity, from probabilistic and empirical models to advanced physical models like Fire Dynamics Simulator (Mell et al., 2006) or FIRETEC. Physical models offer detailed 3D fire simulations but require high-resolution data on vegetation structure, biomass, and fuel properties. Collecting these data, especially in complex Mediterranean landscapes, remains challenging due to ecological heterogeneity and difficult terrain.

Remote sensing has long contributed to vegetation classification and structural characterization, providing key inputs for ecosystem management and wildfire risk assessment (Botella-Martínez and Fernández-Manso, 2017). Satellite or airborne systems have provided detailed data on vegetation composition, structure, and fuel properties. However, limitations such as coarse resolution, cloud interference, high costs, and complex data processing constrain their widespread use. In this sense, Unmanned Aerial Vehicles (UAVs) have emerged as a highly adaptable and high-resolution remote sensing platform, capable of capturing detailed spatial and spectral data across complex and heterogeneous landscapes (Iglhaut et al., 2019). Their increasing integration into forestry research has positioned them as a relevant tool for monitoring and management tasks. UAV-derived aerial photogrammetry (UAV-DAP), particularly when UAVs are equipped with RGB or multispectral sensors, offers a cost-effective way to acquire high-resolution spatial data (Mesas-Carrascosa et al., 2020). These photogrammetric point clouds encode both the geometric structure (e.g., heights, shapes, and spatial distribution of plants) and spectral information (color or multispectral reflectance of vegetation) across the landscape (Yancho et al., 2019; Carbonell-Rivera et al., 2020). However, the lack of dedicated tools and specific methodologies capable of fully leveraging both the geometric and spectral information integrated in the point clouds remains a major limitation to their widespread application in forest environments (Yancho et al., 2019).

This doctoral thesis seeks to fill these knowledge gaps, with the main objective of developing and evaluating methodologies based on UAV photogrammetric point clouds for the analysis of forest structure and fire behavior. Moreover, four specific objectives are described and addressed in the following paragraphs.

The first objective of the thesis “Evaluate the potential of employing spectral, geometric, and neighbourhood features, alongside machine learning methods, for classifying point clouds and identifying vegetation species” was addressed through the design and implementation of a supervised classification methodology tailored to UAV-derived photogrammetric point clouds. The approach involved extracting up to 48 features per point, including geometric descriptors (e.g., planarity, verticality), spectral indices (e.g., NDVI, NGRDI), and neighbourhood-based metrics from RGB and multispectral point clouds. These features were used to train and evaluate five machine learning classifiers: Decision Tree, Extra Trees, Gradient Boosting, Random Forest, and Multi-Layer Perceptron. Fine-tuning of hyperparameters was automated within the workflow to optimize model performance. The methodology was operationalized through the open-source software Class3Dp¹ (Carbonell-Rivera et al., 2024a), specifically developed for the supervised classification of RGB and multispectral (MS) point clouds. In a case study conducted in a Mediterranean shrubland, two classification phases were performed. First, a binary classification between ground and vegetation achieved high overall accuracies: 0.94 with RGB data and 0.95 with multispectral data. Second, points classified as vegetation were further reclassified into five dominant shrub species, achieving overall accuracies of 0.86 (RGB) and 0.87 (MS). Among the classifiers, Gradient Boosting and Extra Trees consistently yielded the highest cross-validation scores. Feature importance analysis highlighted the relevance of verticality and spectral variability within the neighbourhood as key predictors. These results confirm the strong potential of integrating structural and spectral features from UAV-DAP point clouds with supervised machine learning to enable species-level vegetation mapping in complex environments. The Class3Dp software, freely available (Carbonell Rivera et al., 2023), facilitates this process through a user-friendly graphical interface designed to be accessible to users without programming experience, promoting accessibility for users across various fields such as forest ecology, archaeology, and land management.

In the second objective of the thesis “Develop and validate methodologies for processing UAV-DAP data to classify Mediterranean shrub species”, Class3Dp was tested in a real forest environment, applying it to multispectral UAV-derived point clouds collected in two Mediterranean shrubland areas within Sierra Calderona Natural Park, eastern Spain (Carbonell-Rivera et al., 2022). This study explored the potential to classify tree and shrub species based on a combination of spectral, geometric, and neighbourhood features derived from UAV photogrammetric point clouds. Using field-located samples of over a thousand individuals from 11 shrub and one tree species, the methodology involved generating and normalizing dense point clouds, extracting a rich set of features, and training machine learning models to perform semantic segmentation of species at the point level. Gradient Boosting emerged as the most robust classifier, achieving mean cross-validation accuracies of 81.7% and 91.5% for test sites 1 and 2, respectively. After selecting the best classifier, a final segmentation and reclassification step grouped points into individual shrubs or trees, enhancing spatial coherence. When validated with independent field data, overall accuracies of 81.9% and 96.4% were achieved for the two test sites, respectively. The results demonstrated that UAV-DAP, even when acquired with relatively low-cost multispectral sensors, can effectively support species-level classification in complex, shrub-dominated environments. This approach may be particularly valuable for wildfire modeling, as it enables the spatially explicit mapping of species-specific structural attributes that influence fuel characteristics and fire behavior.

The third objective “Investigate the relationships between geometric and spectral variables extracted from UAV-DAP point clouds and the rate of fire spread” enhanced our understanding of how environmental factors, such as vegetation structure and spectral response, influence fire dynamics (Carbonell-Rivera, et al., 2024b). In this study, we utilized UAV-DAP point clouds acquired before the burns across two contrasting environments, open forests and grasslands, where the fire rate of spread (RoS) was mapped using thermal imagery. A grid-based approach was used to spatially align RoS with the geometrical and spectral features derived from RGB and multispectral point clouds and applying Class3Dp. Features included height percentiles, geometric descriptors (e.g., planarity, anisotropy), and vegetation indices (e.g., NBRDI), which were interpolated over optimized grid sizes determined through semivariogram analysis. Regression models, using Random Forest, were applied to explore which combinations of point cloud-derived metrics best explained RoS variability, obtaining R² values of up to 0.56 across the plots studied. Variables such as Planarity_MEAN (planarity mean of the neighborhood, Dist_std_MEAN (the mean of the standard deviation of point distances within the neighborhood), and NBRDI_P75 (75th percentile of the normalized blue–red difference index) consistently exhibited the highest feature importance. Although not all models achieved high predictive power, especially in smaller or less variable plots, results from more extensive and heterogeneous areas demonstrated the potential of UAV-DAP data to capture key fine-scale patterns associated with fire spread. This approach suggests that structural and physiological characteristics of vegetation, derived directly from photogrammetric data, may serve as valuable inputs for enhancing fire behavior models, especially when integrated with topographic and meteorological variables in future studies.

Finally, the fourth and final objective of this thesis was “Estimate aboveground biomass (AGB) of shrub and tree vegetation in Mediterranean ecosystems through the segmentation and classification of UAV-DAP point clouds at individual plant level”. In this work, we implemented a species-specific and individual-level approach by segmenting and classifying vegetation into 13 dominant tree and shrub species across six sites using Class3Dp, while also extracting structural metrics at the individual level to estimate AGB through regression models (Figure 1). Species-level classification achieved an overall accuracy of 81.6%, with values reaching up to 89.9% in the best-performing areas. Regression models for AGB estimation yielded an average R² of 0.69 across all species, with particularly strong results for Anthyllis cytisoides (R²=0.83, RMSE=0.07 kg, n=47), Juniperus oxycedrus (R²=0.83, RMSE=3.17 kg, n=32), and Pinus halepensis (R²=0.77, RMSE=11.79 kg, n=20). These results demonstrate the capability of UAV-DAP to provide detailed and accurate biomass estimates at the individual plant level. Beyond AGB estimation, this study highlighted the potential of UAV-DAP for assessing related ecological indicators such as bulk density and carbon sequestration. The relatively low cost, replicability, and individual-level resolution of UAV-DAP make it an efficient and scalable tool for forest monitoring, especially in Mediterranean-type ecosystems where understory and shrub layers play a critical role in fire dynamics. These findings pave the way for broader use of UAV-DAP in ecological research, climate mitigation strategies, and operational forest management.

These results indicate that UAV-DAP point clouds may hold promise for ecological monitoring and fire modeling in Mediterranean ecosystems. By integrating spectral and geometric features with machine learning algorithms, it was possible to accurately classify Mediterranean shrub species, providing a reliable framework for vegetation mapping. The thesis also revealed that specific geometric and spectral variables derived from UAV-DAP point clouds were correlated with fire spread rates, offering insights into the influence of vegetation structure on fire behavior. Finally, the creation of a methodology for segmentation and classification of individual plants allowed for the successful estimation of AGB of shrub and tree species, underscoring the potential of UAV-DAP technologies to advance our understanding of vegetation structure and support the development of more robust, data-driven strategies for ecological monitoring, wildfire modelling, and sustainable land management in a cost-effective manner.