Assessment of Structural Differences in a Low-Stature Mediterranean-Type Shrubland Using Structure-From-Motion (SfM)

Abstract

Structural traits of vegetation, derived from the three-dimensional distribution of plant elements, are closely linked to ecosystem functions such as productivity and habitat provision. While extensively studied in forest ecosystems, these traits remain understudied in low-stature systems such as Mediterranean-type shrublands. In this study we explore the use of structural metrics derived from small unmanned aerial system (UAS)-based 3D point clouds, generated using the structure-from-motion (SfM) photogrammetry technique, to assess post-fire vegetation structure and biodiversity in the fynbos biome of the Cape Floristic Region (CFR), South Africa. Fynbos is a fire-adapted shrubland that represents nearly 80% of plant species in the CFR, making post-disturbance monitoring critical for conservation. We extracted three structural metrics—canopy height, top rugosity, and surface gap ratio—and achieved ~85% accuracy in classifying 5 × 5 m subplots by burn year using a Multi-Layer Perceptron (MLP), with canopy height as the strongest predictor. Additionally, top rugosity and gap ratio significantly contributed to modeling percentage cover-based species diversity. Our findings demonstrate that UAS-derived structural metrics provide valuable information for characterizing vegetation recovery and biodiversity patterns in low-stature, fire-prone ecosystems. This approach can support ecological monitoring and inform conservation strategies in Mediterranean-type shrublands.

1. Introduction

The Cape Floristic Region (CFR), located at the southern tip of South Africa, is renowned for its rich biodiversity and high level of endemism with species diversity, even competing with highly diverse tropical rainforests [1]. The region covers around 90,000 km² and is home to approximately 9250 plant species, 70% of which are indigenous to the area. The fynbos biome, a Mediterranean-type, fire-prone shrubland, accounts for a significant portion of the plant diversity; in fact, this biome contains approximately 8500 (80%) species, of which around 6000 species are endemic. This makes the region a global hotspot for ecological studies. The high rate of species turnover across fynbos sites and the abundance of locally endemic species accounts for the region’s diversity. This exceptional biodiversity not only makes the region a focal point for ecological research but also supports local economies through ecotourism and conservation efforts. However, these regions are under pressure from several natural and anthropologic impacts. The region has been under constant threat of habitat loss and fragmentation [2,3,4], changing fire patterns [5], exotic species invasion, and a changing climate [6], resulting in the region being designated a Global Extinction Hotspot [7]. In comparison with other highly diverse regions around the world, this region contains one of the highest concentrations of species under worldwide conservation concern [8], of which almost 78% species are endemic to the area [9,10]. It is important to note that the survival of many of these species is highly dependent on prevailing fire regimes.

In fact, fire plays a fundamental and inevitable ecological role in the fynbos biome, which is highly adapted to frequent burning; it has been shown that fire can lead to a great deal of variability in the composition and structure of vegetation, both spatially and temporally [11,12,13]. It is in this context that remote sensing technologies can aid in frequently monitoring and tracking the changes that occur within and between the fynbos sites after they burn, thereby contributing to our ability to estimate the level of biodiversity as a function of space and time. However, it remains a challenging task to detect small changes within the natural environment, where spectral traits (related to how vegetation reflects and absorbs light) and structural traits (related to the physical arrangement of vegetation, such as height and density) do not vary significantly, but still represent high species diversity and functional redundancy [13]. In this context, light detection and ranging (LiDAR) has gained popularity for accurately characterizing and quantifying vegetation structure.

LiDAR systems can provide an accurate 3D structure of the scene under study and have proven to be of significant utility for vegetation modeling and diversity studies, especially when compared to passive multispectral sensors [14,15]. Among the different platforms via which LiDAR data are collected, the use of unmanned aerial systems (UASs) seems best suited to low-stature vegetation modeling, as opposed to other platforms, like satellite or aircraft-based systems [16,17,18]. However, for regions like the CFR fynbos sites, the regular monitoring of a number of burn sites using UAS LiDAR systems is technically and logistically expensive. A more affordable alternative would be the use of structure-from-motion (SfM) approaches, in which point cloud products are derived from high-spatial-resolution imagery. SfM uses images with overlapping regions to construct dense 3D point clouds of a scene by extracting and matching features in the acquired images from multiple perspectives [19]. For a comprehensive overview of SfM techniques and applications, readers are encouraged to consult [20].

SfM-derived data products have gained significant traction in the forestry community for characterizing vegetation structure and linking it to key forest traits [21,22,23,24]. However, there have been few studies dedicated to the use of SfM products for low-stature, Mediterranean-type vegetation. For example, Cunliffe et al. [25] used UAS-acquired SfM data for quantifying the landscape-scale vegetation structure of a semi-arid ecosystem and demonstrated the potential to produce ultra-fine-grain (<1 cm² canopy height models) biophysical products. Abdullah et al. [26] investigated the applicability of UAS image data as complementary 3D data, along with multispectral data, for characterizing desert shrub biomass and to determine factors influencing native vegetation growth in an arid landscape. It was observed that the above-ground biomass within the test sites was correlated with structural traits such as canopy area (R² = 0.81), shrub volume (R² = 0.74), and shrub height (R² = 0.24), and a higher total biomass was associated with higher concentrations of nitrate, phosphorus, soil moisture, and dead organic matter. Similarly, Rivera et al. [27] used UAV-based digital aerial photogrammetry (UAV-DAP) to classify tree and shrub species in Mediterranean forests, achieving classification accuracies of approximately 82% and 96% for two test sites. Their findings demonstrate that accurate species classification using UAV-DAP can serve as a critical step toward extracting structural and fuel load variables, which are essential inputs for improving wildfire behavior models. More recently, van Blerk et al. [28] combined UAV multispectral and SfM data with ground measurements in the CFR to monitor post-fire shrubland recovery and its relation to rainfall seasonality. Their study demonstrated that UAV data can be used to detect fine-scale vegetation changes and complement field observations, although challenges remain in radiometric calibration and species-level identification. Collectively, these studies highlight the broad applicability of low-cost, SfM-derived structural data for assessing vegetation structure across diverse shrubland ecosystems. Importantly, there remains a clear knowledge gap regarding the utility and robustness of SfM data for capturing post-fire structural recovery and diversity patterns, as well as their interrelations, in fire-prone, low-stature shrublands, such as the CFR fynbos. Addressing this gap is essential for developing cost-effective and frequent monitoring approaches to track post-fire structural changes in these ecosystems.

In this study, we use SfM data products to assess structural complexity metrics for low-stature Mediterranean-type shrublands, focusing on fynbos plots of different post-fire ages. Additionally, we aim to quantify structural traits and assess local species diversity (species richness) for each plot, presenting a novel challenge due to the unique ecological characteristics of fynbos. A study by Walter et al. [29] demonstrated a strong correlation between structural complexity metrics derived from terrestrial LiDAR and species richness in the Great Smoky Mountains National Park, USA. Species richness was highly correlated with canopy height (R² = 0.80, p = 0.002) and other depth-related metrics (R² = 0.77–0.84, p ≤ 0.004). Although the study was conducted using terrestrial LiDAR, it highlights the potential of structural complexity metrics for predicting species richness in vegetation ecosystems. More broadly, research in forested ecosystems shows that higher species diversity is often associated with greater structural complexity [30]. This relationship can be attributed to the idea that structurally diverse environments promote species aggregation by offering a wider range of ecological niches [31]. For instance, Deng et al. [32] reported a significant correlation between plant community structure and species diversity, even in areas dominated by herbaceous and shrub layers. Similarly, old-growth structures and canopy gaps were shown to be significantly linked with species richness across multiple taxa in temperate forests in Europe [33]. Internal structural complexity has also been shown to have a stronger relationship with tree species richness and evenness than horizontal or vertical complexity and positively influenced forest productivity across different biomes in China [34]. These studies lay the foundation for the reasoning that structural diversity within a certain ecosystem has a positive correlation with species richness. Although most of these studies focus on forested ecosystems, the underlying principle—that structural diversity fosters species diversity—is likely applicable to other ecosystems as well, including low-stature shrublands such as the fynbos. These insights point to the potential of structural metrics for understanding structural and diversity variations in low-stature shrublands such as the fynbos. In this context, our objectives are (1) to characterize the structural complexity of low-stature vegetation, (2) to differentiate the structures of fynbos burn sites as a function of time-since-last-burn and (3) to model the species diversity based on the structural complexity of fynbos burn sites.

Building on these objectives, we investigate whether structural and species diversity can be modeled using variability in structure and heterogeneity derived from SfM-based metrics. This work is novel in demonstrating how inexpensive SfM-derived structural data, which has not been widely applied to fire-prone Mediterranean-type shrublands, can provide reliable insights into biodiversity dynamics. We posit that greater structural variability is associated with variation in species composition and, therefore, that it can serve as a useful predictor of species diversity. Despite all the physiological phenomena occurring within the complex fynbos biome, our aim is to design a minimalistic and reliable model that can be used for studying and monitoring such ecosystems using structural metrics with an inexpensive 3D dataset, such as one derived from SfM data. We also recognize the limitations of such data and analytical methods—such as potential resolution constraints and the challenge of capturing all ecological nuances—and aim to address these considerations within our study. These challenges underscore the need for cautious interpretation of results and highlight opportunities for future research to refine both data acquisition and modeling approaches to better capture the complexity of such ecosystems.

2. Materials and Methods

2.1. Study Site

The study sites for this work are located in the Grootbos Private Nature Reserve (34°32′S latitude, 19°24′E longitude) near the southern tip of South Africa, as shown in Figure 1. The lower left plot shows six different fynbos sites, which were burnt in the years indicated. A UAS true color image of the 2019 burn plot is shown on the right as an example. Note that the shrubs are unevenly distributed and denser in the middle and southern region than in other parts of the site. Aside from the 2019 burn year, we have five other sites that were burned in the years 2006, 2016 (2 plots), 2020, and 2022, represented by the smaller polygons (lower-left panel in Figure 1); the actual areas covered during our UAS data collection are also shown.

The region is characterized by warm–dry summers and cool–wet winters, typical of Mediterranean ecosystem. It experiences strong seasonality in precipitation, with most areas receiving either winter-dominated rainfall or rainfall that is relatively evenly distributed throughout the year, averaging about 480 mm Mean Annual Precipitation (MAP). The sites are dominated by fire-adapted Fynbos vegetation, with optimal fire cycle between 10 and 14 years. However, fire can be more frequent (2–10 years) due to faster growth rates of many species and the abundance of finer fuel grasses. Natural fires typically occur in late summer and early autumn, towards the end of the dry season, often ignited by lightning or rockfall. The Fynbos biome occupies most of the Cape Fold belt, including its north–south and east–west mountain chains and wetter valleys, as well as adjacent lowlands. The study sites are located within the flat lowlands of the biome. These sites are characterized by nutrient-poor sandy soils that are low in clay content, highly leached, and acidic, posing substantial ecological challenges that drive many of its unique adaptations in the region. Further details on the regional climate, fire regime, and ecological characteristics can be found in [35].

2.2. Data Acquisition

Our team collected imagery for these areas during the month of October 2023 using a DJI Mavic 3M (DJI, Shenzhen, China), a compact and affordable UAS platform, with an RGB camera and a four-band multispectral camera. The RGB images (2.31 cm/pixel from 50 m altitude) were collected following the preplanned flight path, which allowed for the generation of 3D point cloud via structure-from-motion (SfM). The four-band multispectral image frames were not used in this study. The images had approximately 80% along-track overlap and 70% across-track overlap. Pix4Dmapper software (version 4.8.4; Pix4D S.A., Prilly, Switzerland) was utilized for this task to generate georeferenced point cloud data from the collected images and ground-control point geo-information. Readers are referred to https://support.pix4d.com/hc (accessed on 8 August 2025). for detailed information on using Pix4D products for structure generation.

For ground-truthing, our team collected species composition data from six distinct 5 m × 5 m plots in total, one plot each for burn years 2006, 2019, 2020, and 2022, and two plots for burn year 2016. The recorded information within each plot includes the identified species, their abundance, percentage cover, and average height. However, not all variables were consistently available across all burn years due to variations in data availability and field conditions.

Table 1. Description of plot-level ground truth species composition data for different burn plots.

Burn Plots	Nspecies	Dominant Species (>10% Pcover)	Pcover	Abundance	Mean Height (cm)
2006	12	Metalasia muricata	12	18	200
		Passerina corymbose	10	16
		Erica irregularis	30	25
		Indigofera brachystachya	13	12
2016	13	Passerina corymbose	15	16	125
		Erica irregularis	23	22
		Anthospermum aethiopicum	10	25
2016v2	16	Indigofera brachystachya	15	9	65
		Anthospermum aethiopicum	26	45
		Tetraria cuspidate	10	6
2019	10	Restio eleocharis	46	30	40
		Hermannia ternifolia	10	45
2020	12	Restio eleocharis	25	22	30
2022	18	Pelargonium botulinum	20	NA	25

Note: N_species and P_cover represent the number of identified species in the plot and their percentage cover, respectively, while 2016v2 represents the second representative plot for the 2016 burn year.

presents a summary of the collected species composition data for each burn year, including the dominant species and the associated measurements available for each year.

2.3. Data Analysis

To capture structural variability at our study sites, we adopted three structural metrics derived from SfM point cloud data, namely canopy height, top rugosity or external heterogeneity, and surface gap ratio. These metrics, along with internal heterogeneity metric, have been identified by LaRue et al. [36] as better predictors of ecosystem functions such as productivity, energy, and nutrient dynamics. These metrics improve upon traditional biodiversity measures like species richness and phylogenetic diversity, originally derived for dense LiDAR point clouds representing high-stature vegetation. We did not include the internal heterogeneity metric because it is more difficult to estimate reliably with SfM data, particularly in small, dense, and structurally similar shrub-like vegetation, where limited penetration of optical data and occlusion effects hinder resolution of fine-scale internal variation. It is important to reiterate that while the original study used the gap fraction, we replaced it with “surface gap ratio,” which not only includes the entire gap column within a vegetation structure but also accounts for smaller gaps within the canopy surface of the vegetation. A detailed description of the metric extraction process is provided in Section 2.3.1.

The data analysis process, based on the point clouds generated for each burn plot, utilizes spatially explicit maps of the proposed structural metrics. The analysis workflow is shown in Figure 2 (left panel). The UAS-collected imagery over the Fynbos burn sites was processed using structure-from-motion (SfM) techniques to generate high-resolution (approximately 2.3 cm) 3D point clouds.

We derived the digital surface model (DSM), representing the uppermost surface including vegetation and terrain, and the digital terrain model (DTM), representing bare ground, from the SfM-generated point clouds. Their difference produced the canopy height model (CHM), one of the structural metrics used in this study. In addition to the CHM, we calculated two other metrics—top rugosity and surface gap ratio—from the SfM-derived point clouds to characterize vegetation heterogeneity. These metrics were used to quantify the structure of vegetation across different burn years and to assess patterns of post-fire vegetation recovery as a function of time since the last burn. Each burn plot was subdivided into multiple 5 m × 5 m subplots to enable localized structural differentiation. This plot size was chosen based on the standard survey dimensions of a subplot used in this study. Shapefiles containing n subplots per plot were generated using QGIS software (version 3.32.3-Lima; QGIS Development Team, NY, USA). An example for burn year 2019 subplot partitioning is shown in the right panel of Figure 2. The figure shows how the 2019 burn plot was divided into 365 subplots. Depending on the difference in area covered in each flight, the number of subplots for different burn years varied between 192 for burn year 2020 and 505 for year 2006. The three structural metrics were calculated for the n subplots corresponding to each burn year.

Structural metrics were extracted for each subplot and were used to evaluate their ability to discriminate between burn plots based on subplot-level structural variation. Furthermore, these metrics were tested for their association with diversity measures, and used for modeling the relationship between structural complexity and vegetation diversity within the burn plots. A detailed explanation of these analyses is provided in the subsequent sections.

2.3.1. Structural Metrics Extraction

In order to characterize the burn plots structurally to address objective (1) to characterize the structural complexity of low-stature vegetation, we derived the proposed structural metrics (Figure 3) from 3D point cloud data, as follows:

• Canopy Height Model (CHM): To generate the canopy height model (CHM) from the point cloud, 3D points were first classified as ground and non-ground using a cloth-simulation filter (CSF) [37]. The classified points were then rasterized using the Nearest Neighbor interpolation method to create a Digital Terrain Model (DTM) and Digital Surface Model (DSM) for each burn plot. CloudCompare software (version 2.13.2; open-source, EDF R&D, Paris, France) was used to produce these terrain and surface models, and the CHM was obtained by subtracting the DTM from the DSM. The resolution was determined through visual assessment of CSF classification results: when ground points appeared sparse (e.g., earlier burn plots, such as 2006), coarser interpolation (≈1–2 m) was applied for the DTM, even though high-resolution DSMs could be generated. Conversely, when ground points were visually dense and continuous (e.g., recent burn plots, such as 2022), finer interpolation (<0.1 m) was used for the DTM; however, these plots typically had fewer non-ground points, so a correspondingly lower resolution was selected for the DSM. Thus, the choice of resolution in creating DSM and DTM was based on the analysis of quality of ground points classification.
• Top Rugosity or External Heterogeneity (TR): Top rugosity was calculated by running a small window (3 × 3) over the CHM and calculating the standard deviation of canopy heights within the window, which basically represents how the surface of the canopy varies spatially within the plot. This metric represents within-plot top surface variability.
• Surface Gap Ratio (GR): The surface gap ratio, or simply gap ratio, was generated via a similar window operation to top rugosity, by determining the ratio of number of points in the window column with heights lower than the mean height of points within the window, ratioed by the total number of points in the window. For the window column, which mostly contains ground points, this metric will be lower (ideally near 0), and for dense surfaces, it will be higher, depending on the number and density of points within the window column. Thus, this metric at each pixel represents the proportion of gaps at the surface of study plots.

A consistent window size of 3 × 3 and grid resolution of 0.25 m were used to determine the gap ratio metrics. The grid resolution was chosen to be 0.25 m because it was found to be optimal for determining this metric in terms of computational time and memory requirements on standard desktop workstation (approximately 16–32 GB RAM), as well as sampling accuracy, which is also shown in Section 3.6 of this document. The geolocated maps for each plot and metric were obtained by performing these operations on the SfM-derived point clouds. From this point forward, these structural metrics will be referred to as ‘CH’ or ‘CHM’ (canopy height), ‘TR’ (top rugosity), and ‘GR’ (gap ratio).

2.3.2. Subplot Supervised Classification

In accordance with objective (2) to assess the ability of the proposed structural metrics in classifying the subplots into burn year, we tested six frequently used supervised classification algorithms:

• Gaussian Naive Bayes (GNB): The Naive Bayes classifier [38] is based on the Bayes theorem of conditional probability, which assumes that (1) the features of data are conditionally independent from each other for a given class label and (2) all features have equal importance in predicting the class label. With these assumptions, the general Bayes theorem reduces to the Naïve Bayes model, given by

$$P\left(y|X\right)\propto P\left(y\right)\prod _i^{n}P\left(x_i|y\right)$$

(1)

where $P\left(y\right|X)$ is the posterior probability of class y, given X, P(y) is the prior probability of class y, $P\left(y\right)$ is the probability of feature $x_i$, given class y, and n is the number of features in the feature vector X. We therefore need to find the probability of a given set of input values for all possible values of the class variable y and select the one with maximum probability, in order to create a classifier model with Naive Bayes. The formulation for decision rule for the classifier thus reduces to

$$\widehat{y}=\mathit{arg}\underset{y}{\mathit{max}} P\left(y\right)\prod _i^{n}P\left(x_i|y\right)$$

(2)

where $P\left(y\right)$ is a conditional probability represented by Gaussian distribution.

2.

Support Vector Machine (SVM): Support Vector Machine (SVM) [39] is another popular supervised machine learning algorithm, which finds an optimal hyperplane in an N-dimensional space that can separate the data points into different classes in the feature space. SVMs, by nature, are binary classifiers, but to use them for a multi-class classification problem, we used a one-vs.-one decision rule, meaning that a separate classifier is trained for every pair of classes and when making a prediction, each classifier predicts a class for the input, and the class that is predicted the most frequently is selected as the output label. The decision rule for binary SVM classifier is given by

$$y_i=\left\{\begin{array}{c}1 if w^T{x}_i+b\ge 0\\ 0 if w^T{x}_i+b<0\end{array}\right.$$

(3)

$$Minimize:{\displaystyle \frac{1}{2}}{w}^{T}w+C\sum _{i=1}^N{\xi }_i$$

$$Subject to: y_i\left(w^T\phi \left(x_i\right)+b\right)\ge 1-{\xi }_i,i=1,\dots ,n$$

$${\xi }_i\ge 0,i=1,\dots ,n$$

where the following apply:

-

w is the weight vector;

-

C is the regularization parameter;

-

φ(x) is the feature map;

-

$x_i$ is the i-th training sample;

-

$y_i$ is the corresponding class label;

-

${\mathsf{\xi }}_{\mathrm{i}}$ are slack variables;

-

b is the bias term

3.

K-Nearest Neighbor (KNN): KNN is a non-parametric supervised learning algorithm that uses a proximity determination to make classification [40]. An input is classified by the number of votes of its neighbors, and it is classified into the class that is most common among its k-nearest neighbors. The distance metric used for determining the nearest neighbors is Euclidean distance.

$$D\left(X,Y\right)={\left(\sum _{i=1}^n{\left|x_i-y_i\right|}^p\right)}^{\left(1/p\right)},p=2$$

(4)

4.

Decision Tree: A Decision Tree classifier operates by creating a tree-like structure, where each internal node represents features; specifically, branches represent the decision rule, and the leaf nodes represent the output class label [41]. The tree is constructed by repeatedly splitting the training data into smaller subsets, based on some metric/s, until a stopping criterion is met. Metrics such as entropy (measure of randomness) or Gini impurity (level of impurity) are used to find the best attribute to split the dataset.

$$H\left(t\right)=-\sum _{i=1}^{k}p\left(i|t\right)lo{g}_2\left(p\left(i|t\right)\right)$$

(5)

$$Gini\left(t\right)=1-\sum _{i=1}^k{\left(p\left(i|t\right)\right)}^2$$

(6)

H(t) is the entropy metric and Gini(t) is the Gini impurity metric. The goal is to determine the feature and split the dataset with regard to the feature that results in the greatest reduction in impurity, leading to more homogeneous “child nodes” (internal nodes). This process continues recursively until we reach a level where we cannot split a node anymore, or we reach a maximum depth tree. We then navigate through the tree testing at each node, moving to the left or right of the node based on the test result, in order to make predictions on new instances. The final leaf node where we finish will be assigned as the label for the instance.

5.

Random Forest: The Random Forest method [42] operates by making decisions based on predictions of ensembles of decision trees. These trees are formed by generating a number of bootstrapped datasets (random subset of training data with replacement) from the original dataset and selecting random features for each bootstrapped dataset to produce the tree. The randomness in feature selection helps in preventing overfitting during the learning process. The prediction for the new sample is made on the basis of the majority of votes given by each decision tree. Random Forest can model complex decision boundaries and does not require scaling of input features.

6.

XGBoost: The XGBoost method [43] builds an ensemble of decision trees in a step-by-step manner, with each new tree focusing on correcting the errors made by the previous ones. It uses gradient boosting, a process in which trees are added one at a time and each tree fixes the mistakes of the previous ones to optimize predictions, and it includes built-in regularization. This method helps to prevent overfitting in the model training process. The predictions for new samples are made by combining the weighted contributions of all trees in the model. XGBoost is known for being fast and efficient, works well with large datasets, and provides high predictive accuracy.

7.

Multi-Layer Perceptron (MLP): Multi-Layer Perceptron (MLP) is a type of feedforward artificial neural network capable of modeling complex nonlinear relationships, often used for supervised classification [44]. The input features were standardized (zero mean, unit variance), and class labels were one-hot-encoded. The network architecture included three hidden layers with 64, 128, and 64 neurons, respectively, using ReLU activation function, L2 regularization (0.01, 0.01, 0.005), batch normalization, and dropout (0.3, 0.4, 0.3). The output layer used softmax activation with neurons equal to the number of classes. The model was trained using the Adam optimizer and categorical cross-entropy loss, with early stopping (patience = 15) and adaptive learning rate reduction (factor = 0.5, patience = 5, min_lr = 1 × 10⁻⁶). Training was performed for up to 100 epochs with a batch size of 8 while 20% of the data was used for validation.

As noted in previous section, the number of partitioned subplots varied between burn years and resulted in class imbalance during the training of the classifier, i.e., the classifier “sees” some classes more frequently (majority class) than others (minority class), thus resulting in bias during the training. More specifically, the samples from the burn year 2020 represent the minority class (192) and samples from the burn year 2006 are the majority class (505). We used SMOTE (Synthetic Minority Oversampling Technique) [45], to generate samples for the minority class without changing overall class statistics, in order to mitigate this problem. We first divided our original data into a train-test split of 80–20% samples and then used SMOTE to resample the training data, keeping the test data the same. This approach was taken because we want to use the generated synthetic instances only while training, and while testing, we want to use the samples from unmodified distribution of original test data, which does not contain any synthetically generated data. All the burn year classes had an equal number of samples for training the classifiers after applying SMOTE and the total number of samples were scaled from 1695 to 2525 (505 per class). We used a t-test and visual analysis of class statistics before and after applying SMOTE (Figure 4) to ensure that we were not compromising the integrity of original data while oversampling. In this way, we aimed to balance the class distribution while preserving the overall structure and variability of the original dataset. Each model’s performance was evaluated on the test set containing a total of 428 samples. A classification report comprising test accuracy, precision, recall, and confusion matrix was generated for further evaluation.

Furthermore, the feature importance was computed to evaluate the relative contribution of each structural metric in the classification task. For classifiers that provide inherent feature importance scores (e.g., Random Forest, XGBoost, and Decision Tree), the built-in importance measures were directly used. For classifiers that do not natively offer feature importance (e.g., MLP, SVM, Gaussian Naïve Bayes, and KNN), permutation-based importance was computed following the implementation in scikit-learn [46], which is based on the original idea by Brieman [47]. In this approach, the values of each feature were randomly permuted across the test dataset while keeping all other features unchanged, and the resulting decrease in model performance was recorded. A greater drop in performance indicated a higher importance of the permuted feature. This strategy enabled a consistent and comparable estimation of feature importance across all classifiers. It is to be noted that our goal is not to propose a new method or algorithm for classification, but rather to leverage the existing methods in fulfilling our objective of classifying the subplots into their corresponding burn year and determine the contribution of each structural metric in such a classification.

2.3.3. Diversity Modeling

With regard to objective (3) to model the species diversity based on the structural complexity of fynbos burn sites, we propose that the variability in the percentage cover-based diversity index, namely Shannon–Wiener Index (H’) or Shannon Entropy [48], can be modeled using SfM derived structural metrics. Percentage cover data quantifies the relative dominance of species within a sampling unit. In fact, this index measures alpha-diversity and represents local structural diversity within burn plots [49]. For any plot, given percentage cover information, the Shannon–Wiener Index can be calculated as

$$H^{\prime }=-\sum _{i=1}^N{p}_i\cdot \ln \left(p_i\right)$$

(7)

where the following apply:

$p_i= {\displaystyle \frac{cove{r}_i}{{\sum }_{j=1}^{S}cove{r}_j}}$ is the proportion cover of each identified species;

$N$ is the total number of species identified.

The Shannon index typically ranges from 0 (single species dominance) to higher value H_max, depending upon the ecosystem diversity. We used percentage cover-based calculations, as our goal was to relate structural metrics to species diversity by capturing the spatial patterns of aggregation and dominance. Compared to simple species counts, percentage cover aligns more closely with the structural characteristics derived from SfM data, making it a more appropriate measure for evaluating diversity in this context. Furthermore, abundance- and species count-based diversity indices have been shown to correlate closely with cover-based indices [50,51], suggesting that models developed using cover-based diversity are likely to be broadly applicable to other commonly used diversity metrics as well.

We computed the average values of the three structural metrics for each burn plot: canopy height (CH), top rugosity (TR), and gap ratio (GR). The percentage cover data obtained from a corresponding reference field plot within each burn plot was used as the representative diversity measurement for that burn plot, enabling a one-to-one pairing between structural metrics and diversity indices. To model the relationship between structural complexity and diversity, we formulated weighted combinations of the normalized structural metrics (normalized across burn year). Given the limited number of available samples, we employed an exhaustive grid search to explore all possible combinations of weights (in increments of 0.05) for TR, GR, and CH, constrained to sum to 1. For each weight combination, a composite structural index was computed and evaluated through linear regression against the corresponding Shannon–Wiener diversity index. The coefficient of determination (R²) was used to assess the strength of association, and the weight combination yielding the highest R² value was selected as the optimal representation of structural contribution to diversity.

3. Results and Discussion

3.1. Extraction and Mapping of Structural Metrics—Observations from a Visual Analysis

The structural metric maps for all six burn plots were generated, with the maps for the 2019 burn year presented in Figure 5 as a representative example. The canopy height and top rugosity metrics were produced at relatively high resolutions compared to the gap ratio metric, which was generated at a fixed resolution of 25 cm. The resolution of the canopy height model and its derived products—such as top rugosity—varied across plots and was influenced by the effectiveness of the cloth-simulation filter in classifying ground versus non-ground points. In earlier burn plots (e.g., 2006), surface points dominated the classification process, reducing ground point detection, whereas in more recent plots (e.g., 2022), ground points were more reliably identified. Based on visual inspection and evaluation of the point classification quality, the most appropriate grid resolution was selected for each burn plot.

The gap ratio maps, on the other hand, were derived directly from the point cloud using the Python programming language (version 3.10.11; Python Software Foundation, Wilmington, DE, USA). Generating these at resolutions finer than 0.25 m proved computationally intensive and yielded little improvement in output quality. Furthermore, a sensitivity analysis on varying grid resolutions (see Section 3.5) indicated minimal impact on the final results. Therefore, a 0.25 m resolution was adopted for all gap ratio maps in subsequent analyses.

Figure 5 shows that most of the area is covered by small shrubs for the burn year 2019. The areas with taller and denser shrub species, i.e., locations with significant growth after the 2019 burn, are highlighted by each metric in its respective map. The denser species have an average height of ~1.5 m, even though the plot is mostly dominated by smaller species with lower heights (<0.5 m). It is evident from the top rugosity map that the 3D surface is changing: most of the site is dominated by small species that do not exhibit much change in surface height, with only the regions where the denser shrubs are present showing variability in height. Similarly, the gap ratio map shows larger gaps in regions where there are no dense shrubs present. This example figure shows how these structural metrics can distinguish different structural features in the fynbos as a function of burn year, which we will analyze in the next section to differentiate subplots by burn year.

3.2. Subplot Quantitative Analysis

The distribution of structural metrics across all the burn year sites is shown in Figure 6. In the figure, the ‘canopy height’ distribution shows that the vegetation in the earliest burn plot (2006) attained heights of ~2 m, forming a dense, uniform canopy characteristic of late-successional shrublands, while the species in the most recent burn plot (2022) is just starting to re-establish, with an average height of ~0.2 m. The ‘top rugosity’ distribution, in turn, shows that the plots that were burnt early (2006) and recently (2019, 2020, and 2022) possess lower rugosity, on average, when compared to the mid-burn years (2016). This was attributed to the fact that in the early burn years, almost all shrubs had reached their peak growth and existed in a dense configuration; the difference in their height, therefore, is limited, while the recent burn plots are just starting to re-establish successional species. The majority of the surface in the latter case, therefore, is covered by either ground or short-stature/young species, leading to small height differences, but at lower heights. In contrast, the maximum height variation is observed around 2016, representing a mid-successional stage with mixed species composition and varying growth stages contributing to higher structural complexity. The ‘gap ratio’ distribution shows that the gap ratios in burn year 2006 are the lowest, while burn years from 2016 onward exhibit similar gap ratios, with burn year 2019 showing a slightly higher mean gap ratio. Overall, the trend shows that as a function of time since burn, canopy height steadily increases as vegetation regrows; top rugosity initially increases, reaches a peak at intermediate stages, and then declines as the canopy becomes more uniform; while gap ratio consistently decreases over time, reflecting a reduction in canopy gaps as the vegetation matures.

The structural metrics were reduced to a single composite variable using Principal Component Analysis (PCA) [52] in order to assess the similarity between the structures of the burn plot samples; this PCA component explained nearly 99% of the variance in the original data. Thus, the use of this composite metric preserved most of the information contained in the original three metrics, while (i) significantly reducing data dimensionality and (ii) reprojecting the data onto the principal axis that best explains the variance. The distributions of PC1 scores for each burn plot are visualized as violin plots in Figure 7, providing a compact summary of the structural variability across different burn years.

The figure reveals that the samples from the 2006 burn plot are distinctly separated from the other groups, reflecting their mature and structurally divergent characteristics relative to the younger burn plots. In contrast, substantial overlap is observed among the 2016 and 2019 samples, as well as among the 2019, 2020, and 2022 samples, suggesting considerable structural similarity within these groups. This overlap implies that, during classification, samples from 2006 are likely to be the most distinguishable due to their unique structure, whereas samples from overlapping burn years may exhibit higher classification uncertainty and potential misclassification due to their structural resemblance.

3.3. Subplot Classification

The results of classifying subplots into burn year classes based on the structural metrics are presented in

Table 2. Classification performance of different machine learning models for burn year prediction on the independent test dataset (N = 428). The table shows class-wise F1-scores for each burn year and overall test accuracy.

Classifier	Burn Year Prediction on Test Data [F1-Scores]					Overall Test Accuracy (%)
	2006	2016	2019	2020	2022
MLP	1.00	0.84	0.65	0.74	0.81	84
Random Forest	1.00	0.82	0.67	0.74	0.78	83
SVM	1.00	0.83	0.64	0.71	0.82	83
XGBoost	1.00	0.83	0.66	0.70	0.82	83
Naïve Bayes	1.00	0.83	0.59	0.72	0.76	81
KNN	1.00	0.80	0.68	0.63	0.73	80
Des-Tree	1.00	0.76	0.59	0.64	0.64	76

. Each of the trained classifiers was tested on a total of 428 new, unseen samples, representing 20% of the data from each class combined. This approach ensured that none of the six burn year classes were omitted during testing and prevented over-representation of certain classes at the expense of others. The F1-scores for each class are reported as the performance metric, which is defined as the harmonic mean of precision (User’s accuracy) and recall (Producer’s accuracy), providing a balanced assessment of the model’s classification performance; higher scores indicate better performance. Among the six classification algorithms, MLP yielded the highest prediction accuracy, of approximately 84%, while Decision Tree resulted in the lowest accuracy, of approximately 76%.

We also observed that the subplots from the burn year 2006 were classified perfectly by all the classifiers and were the easiest samples to predict. This is logically consistent, as burn year 2006 was structurally distinct from the other burn plots based on the distribution of all the structural metrics. In contrast, burn year 2019 was the most challenging class to predict, with an F1-score as low as 65% for the MLP classifier. The prediction for each class with MLP is shown in Figure 8A in a confusion matrix. From the figure, we can see that most of the misclassification instances occurred between burn years 2016 and 2019, as well as in 2019, 2020, and 2022, reflecting the structural similarities shared by these plots. The classifiers tended to confuse samples with similar growth stages, yet still achieved the best prediction accuracy, of approximately 84%. The right panel in Figure 8 presents the contribution of each metric across all classifiers in producing the classification outcomes. Across all algorithms, canopy height consistently emerged as the most important feature for classifying subplots. The top rugosity (TR) and gap fraction ratio (GR) metrics also contributed meaningfully to the classification, although their relative importance varied across algorithms. Specifically, GR was the second most important feature for the Decision Tree, MLP, and KNN classifiers, whereas TR was ranked second for Gaussian Naïve Bayes and SVM. For Random Forest and XGBoost, both TR and GR demonstrated comparable importance, contributing similarly to the model’s predictive performance.

Overall, these results suggest that while canopy height serves as a primary structural indicator of post-burn recovery stage, incorporating additional structural metrics such as rugosity and gap fraction ratio allows the models to capture more subtle differences among burn years. This highlights the value of using a combination of structural descriptors to improve classification performance, particularly when distinguishing between burn years with overlapping structural characteristics.

3.4. Diversity Modeling Outcomes

The diversity modeling results are summarized in Figure 9. Based on the five burn year test samples (2006, 2016 [two plots], 2020, and 2022), the optimal model assigned equal weights of 0.5 to both top rugosity (TR) and gap ratio (GR), while canopy height (CH) did not contribute to the final model. The 2019 plot was excluded from model development due to its substantially lower diversity relative to both earlier (2006, 2016) and later burns (2020, 2022), indicating it was not representative of the broader post-burn recovery pattern. The optimized model demonstrated a strong linear relationship between the composite structural index and the observed Shannon–Wiener diversity index, with an R² of 0.86 (p = 0.023) and a mean absolute error (MAE) of 0.03 (Figure 9B). The dominant contribution of TR and GR, with no contribution from CH, suggests that horizontal and fine-scale vertical heterogeneity play a stronger role in explaining the diversity across the sampled plots than mean canopy height. In low-stature vegetation systems such as Fynbos, in recently disturbed ecosystems (e.g., post-burn recovery), small-scale structural variability (captured by TR and GR) may better reflect microsite availability, niche heterogeneity, and light penetration—all key drivers of plant community composition. In contrast, CH represents an average vertical structure that may not capture these finer-scale habitat differences as effectively, especially when the overall height differences across plots are relatively small. Additionally, because structural metrics were aggregated at the plot level, mean-based measures such as canopy height may overlook ecologically significant fine-scale variability. In contrast, heterogeneity metrics like top rugosity and gap ratio are more sensitive to within-plot structural complexity and may better capture localized microsites that support plant diversity.

These findings align with prior research highlighting the effectiveness of structure-derived metrics, especially those reflecting spatial heterogeneity, in predicting species diversity [29,53,54]. However, given the limited number of samples and the use of plot-level averages, caution is warranted when generalizing these results. Additional sampling, particularly across a wider range of structural conditions and disturbance histories, may help further elucidate the relative contributions of different structural components to diversity.

Figure 10 presents the estimated diversity maps for three representative burn years, derived using the developed structural metric-based model. Each map consists of 5 × 5 m subplots represented by each pixel, with the color intensity reflecting the predicted Shannon–Wiener diversity index. The subplots indicated by yellow pixels exhibit higher diversity values, while the dark blue pixels represent areas with lower diversity. The low-diversity patches (dark blue) generally correspond to regions dominated by shrub monocultures or structurally uniform species, where individuals exhibit similar growth patterns following fire disturbance. This is expected, as structurally similar species tend to regenerate and grow at comparable rates post-burn, reducing heterogeneity in structural traits and thus limiting diversity—a pattern widely observed in fire-prone shrub communities [55].

Across the different burn years, distinct spatial and temporal patterns are observed. The 2006 plot, representing the earliest burn, shows relatively uniform diversity across most of the area, consistent with its more mature successional stage, where dominant species have reached full growth and structural homogenization is higher. For the 2016 plot, which represents a mid-successional stage, greater spatial variability is evident, particularly in the central region, where taller, denser shrub formations are observed. Field observations confirm that this central zone is characterized by denser, structurally complex vegetation, contributing to higher diversity estimates. In contrast, peripheral areas contain shorter or less mature shrubs, exhibiting lower structural heterogeneity and, thus, reduced diversity. The 2019 plot, representing the recent burn, displays the highest spatial variability in estimated diversity. This variation likely reflects early successional dynamics, where multiple species at different stages of recovery coexist, resulting in heterogeneous growth forms and niche partitioning across the plot.

A KDE plot of diversity estimates for all subplots across burn years is shown in Figure 11. Each curve represents the distribution of subplot-level diversity estimates for a given burn year, with the vertical stacking aiding comparison across years. To visualize the temporal trend, the mean diversity value for each burn year was computed and connected by a dashed line. The overall trend observed from the KDE distributions shows that diversity is lowest for the earliest burn year (2006), increases and peaks around the 2016 burn years (both 2016 and 2016v2), and subsequently declines for the more recent burns (2019 and 2020), with a slight increase again for 2022. This pattern reflects both successional dynamics and the influence of fire history, with diversity initially increasing following disturbance as multiple species establish, but then declining over time as competitive exclusion and resource limitations reduce community heterogeneity. Such patterns are consistent with prior ecological observations in fynbos systems. For example, Verboom et al. [56] report that species richness tends to decline with increasing time since fire, partly due to nutrient depletion as the post-fire ash layer dissipates. These results demonstrate that the proposed structural metric model is capable of capturing both temporal and spatial patterns of diversity following fire, providing biologically consistent estimates across different successional stages.

3.5. Effects of Vegetation Structure on Percentage Cover vs. Abundance-Based Diversity

As a subsequent analysis, we examined how the three structural complexity metrics—canopy height (CH), top rugosity (TR), and gap ratio (GR)—relate to species diversity under two formulations: one based on percentage cover, and another based on species abundance. The analysis was considered in order to determine whether different aspects of structural complexity are more or less predictive, depending on the type of ecological diversity being assessed. However, it is important to note that the available data was limited in sample size and temporal coverage. The abundance-based diversity index was available for four burn years—2006, 2016 (two plots), and 2020—while the percentage cover-based index was available for five burn years, including an additional plot from 2022. Therefore, the observed patterns should be interpreted cautiously, particularly for the abundance-based metrics. Figure 12 illustrates the relationship between percentage cover and species abundance for the four burn years where both data were available: 2006, 2016 (two plots), and 2020.

In the earlier burn plots (e.g., 2006 and 2016), the relationship between percentage cover and abundance appears to be nearly linear, indicating a relatively even distribution of species. In contrast, the 2020 plot shows a strong deviation from this pattern, with one species dominating both in percentage cover and abundance. This suggests that the more recent burn plots are often characterized by early successional dynamics, in which a few fast-growing species dominate, while the older plots tend to exhibit greater species evenness. These patterns reflect the typical trajectory of post-disturbance recovery, with vegetation composition becoming more balanced as structural complexity and microsite heterogeneity increase over time. Moreover, this variation also highlights how percentage cover-based and abundance-based indices can yield different interpretations of species diversity, particularly in recovering ecosystems.

When diversity was calculated using the percentage cover-based Shannon index, TR showed a moderately positive correlation (Pearson r = 0.65, p = 0.235; Spearman ρ = 0.50, p = 0.391), suggesting that greater rugosity—representing variability in vertical structure—may promote a more diverse range of vegetation types at the surface level. GR also exhibited a weak positive correlation (Pearson r = 0.26), potentially reflecting the influence of canopy openness in supporting varied vegetation cover. CHM, by contrast, showed a weak negative relationship (Pearson r = –0.29), indicating that mean vegetation height alone may not strongly influence surface-level compositional diversity. However, as explained above, only TR and GR entered the model for predicting diversity.

In contrast, the abundance-based Shannon diversity index revealed different associations. TR demonstrated a strong negative correlation (Pearson r = –0.89, p = 0.108), and CH also showed a more substantial negative trend (Pearson r = –0.68, p = 0.320). GR, on the other hand, maintained a moderate positive correlation (Pearson r = 0.64, p = 0.358). These results suggest that plots with greater top rugosity canopies may be structurally dominated by few species, reducing evenness and, hence, lowering abundance-based diversity. Conversely, a higher gap ratio promotes greater species coexistence through increased light penetration and microsite heterogeneity, supporting both richness and evenness in species composition.

Overall, these findings indicate that the choice of diversity index considerably influences which structural metrics appear ecologically relevant. Percentage cover-based indices capture surface-level diversity patterns, such as the spread of vegetation across the ground, while abundance-based indices reflect vertical structure by highlighting dominant, taller species and their competitive influence. These trends are consistent with previous findings that highlight the complex and scale-dependent ways in which vegetation structure influences plant diversity patterns [57,58]. The observed differences also underscore the need for care when interpreting trends from limited temporal samples. In particular, the strong correlations seen in the abundance-based analysis may be overrepresented due to the small number of burn years considered. Future work with more temporally diverse sampling will be necessary to validate and generalize these relationships in a more robust approach.

3.6. Impact of Sampling Resolution

Our final analysis evaluated the effect of grid spatial resolution on the computation of the ‘gap ratio’ (GR) metric. This was conducted to determine a suitable interpolated grid resolution for calculating structural metrics from point cloud data. The GR metric was computed at multiple grid resolutions: 0.125 m, 0.25 m, 0.5 m, 1 m, 2 m, 4 m, 8 m, and 10 m. The results indicated that the impact of resolution on the GR metric was minimal. The mean GR values of each subplot across all burn years and tested resolutions are shown in Figure 13. Noticeable changes in the mean metric values emerge only beyond the 4 m resolution. However, the differences in magnitude are not substantial, which is likely due to the overall structural uniformity within most of the burn plots, with limited areas of high variability. Such variability becomes more apparent at finer grid resolutions or when integrating the metric over larger field areas.

Quantitatively, the RMSE between GR values calculated at the 0.125 m and 4 m resolutions was approximately zero, indicating that these resolutions yield nearly identical metric estimates. At coarser scales, the RMSE values between 0.125 m and 10 m, and between 4 m and 10 m, remained relatively small (0.04 for 2006 and 0.07 for 2019), suggesting that noticeable differences in metric values begin to appear primarily when the grid resolution exceeds 4 m. These results demonstrate that the variation in calculated GR values remains minimal at resolutions finer than 4 m and, thus, the choice of grid resolution below this threshold is unlikely to affect metric computation or subsequent analyses of structural traits at this spatial scale.

3.7. Operational Feasibility

The operational feasibility is a key consideration for implementing UAS-SfM-based post-fire monitoring in low-stature shrublands. UAS-based platforms are affordable and flexible, and they can be flown on demand without scheduling delays. Flight parameters such as altitude and image overlap determine the data volume: lower flight altitudes (finer GSD) yields finer details, but more images are captured to cover an area compared to higher flights. In practice, mapping a 1 ha area with height overlap (>80%) can require hundreds to thousands of images and 5–15 min of flight, whereas flying higher drastically reduces the image number count and flight duration, while producing coarser-resolution point clouds [59].

Our study utilized an off-the-shelf four-band DJI Mavic 3M (US$3000–4000) system, equipped with consumer-grade cameras, generating point clouds at 2–3 cm GSD. The survey area comprised ~2 ha fynbos burn plots; each survey required approximately 5.5 min of flight time and generated about 3 GB of imagery, of which roughly 20% corresponded to the RGB bands alone. SfM processing in Pix4D mapper was completed in ~15–20 min on a high-end workstation, 512 GB RAM, GPU:NVIDIATITANV (Driver: 31.0.15.1601), CPU: Intel(R) Xeon(R) CPUE5-2650 v4 @2.20 GHz, Windows Server 2019 Standard, 64-bit, and yielded point densities exceeding 500 points/m³.

Affordable consumer UASs (1–2 kg payload) with multispectral cameras cost a few thousand dollars in total, whereas higher-end rigs (e.g., UAS-LiDAR) can be roughly 10× more expensive [60,61]. The main recurring cost is personnel time: field campaigns are quick, although specialized processing/analysis skills are required. A cost analysis reported that UAS fieldwork (flying + GCPs) with a DJI Mavic 3M took ~25 h for 8 km² (800 ha) at 120 m flight height and 10 cm GSD [62]; scaled down to our sites, this equates roughly 1–2 h per site. Recurring costs include rechargeable batteries (~USD 50–100 each) with 5–10 year lifespans, along with minor maintenance and insurance costs.

As the monitoring frequency increases, UAS-based approaches become more cost-effective than manual surveys [63]. Annual or semi-annual flyovers are practical, since the marginal cost per flight is low (primarily battery and maintenance), and the fine spatial resolution directly captures recovery dynamics. With declining equipment costs and increasing automation of workflows, UAS-SfM monitoring is expected to become even more economical than traditional field assessments. For small areas (~1–2 ha), data collection takes only minutes and requires modest labor and equipment. SfM processing is computationally intensive; however, modern workstations and increasingly efficient open-source software help mitigate these demands. Overall, UAV–SfM provides an operationally feasible and high-resolution approach for the routine monitoring of post-fire recovery in fire-prone shrublands.

4. Conclusions

We presented a method to characterize structural differences and diversity, as a function of burn year, in a low-stature fynbos (Mediterranean) shrubland ecosystem using inexpensive structure-from-motion (SfM) point cloud data obtained via an off-the-shelf UAS. Three structural metrics—canopy height, top rugosity, and surface gap ratio—representing multiple aspects of vegetation structure, were proposed for this task. We first performed a plot-level comparison of these metrics across five study sites that had burned in the years 2006, 2016, 2019, 2020, and 2022. These plots were then partitioned into smaller subplots (5 × 5 m), and the three metrics were computed at the subplot level. Seven classification algorithms were evaluated to assign subplots to burn year classes, with the Multi-Layer Perceptron (MLP) achieving the highest classification accuracy of approximately 85%. This finding has important implications for conservation of the fynbos biome. In short, data acquisition with an affordable, four-band UAS can capture multi-perspective imagery, followed by SfM processing to derive 3D point clouds, enabling the classification of burn year as a function of three structural traits. This approach, in turn, allows for accurate assessment of plot-level species and structural diversity. Such outputs provide researchers and ecologists with valuable insights into biodiversity patterns within this low-stature, fire-prone, successionally dynamic ecosystem, and offer a practical and conservation-relevant alternative to more expensive and computationally intensive LiDAR systems. This also highlights the novelty of applying simple SfM-derived metrics to reliably capture post-fire structural and biodiversity patterns in this biome, where such monitoring has been largely unexplored.

In addition, we developed a model to estimate diversity based on top rugosity and gap ratio metrics, achieving an R² of 86% and a MAE of 0.03, and produced associated landscape-level diversity maps across several burn years. These results demonstrate strong potential for assessing and mapping species diversity variation both within and across plots, and represent a significant advance by linking structural recovery metrics to biodiversity outcomes. Such integration provides actionable information for restoration planning, prioritizing areas for intervention, and tracking the ecological resilience of Mediterranean-type shrublands.

However, several aspects of this analysis warrant further investigation. The reference subplots were assumed to be representative of all the burn plots, and most of the results were validated using limited ground truth data and in situ visual assessments by trained botanists and ecologists. Future studies should explore scaling such plot-level estimates by incorporating additional subplot-level truth data. We also suggest expanding the set of structural metrics to include additional traits, as well as spectrally derived trait measurements (e.g., nitrogen content from red and near-infrared reflectance), as three structural metrics may be limited in fully capturing trait complexity in such hyper-diverse ecosystems. It is also important to recognize that the goal of this study was to assess structural diversity in the fynbos biome as a function of time-since-burn, and that the relationship between structural variation and diversity indices may vary under certain ecological conditions. Furthermore, additional quantitative validation with a larger truth dataset would enable improved error analysis when generating diversity maps.

Overall, we conclude that the combination of burn year classification and diversity estimation using relatively simple SfM point cloud datasets offers a practical tool for mapping, monitoring, management, and conservation of Mediterranean low-stature ecosystems. This is particularly critical for the fynbos biome, which exhibits exceptionally high species diversity and endemism.