Cork Oak Regeneration Prediction Through Multilayer Perceptron Architectures

Abstract

In Mediterranean ecosystems, a thorough understanding of seedling regeneration dynamics as well as a good predictive ability of the process is essential for sustainable forest management. Leveraging the predictive capacity of the multilayer perceptron (MLP) as recognized as artificial intelligence methodology, the authors analyzed a real case study with a dataset encompassing environmental, ecological, and forestry variables. The study focused on the cork oak (Quercus suber, L.) seedling regeneration dynamic, which is a critical process for maintaining ecosystem resilience. A set of 10 MLP with a block from 5 to 50 neurons with hyperbolic tangent (TanH), linear (LIN), and Gaussian (GAUS) activation function were tested and their performance for predictive purposes was compared with traditional quantitative approaches. The MLP configured with 40–50 neurons per activation function (TanH, LIN, GAUS) demonstrated outstanding predictive performance, achieving an area under the curve (AUC) of the receiver operating characteristic and precision-recall scores above 0.80. These models made few prediction errors, effectively explaining the majority of the data variance, as indicated by a high generalized R² and a low mislearning ratio. This approach outperformed traditional statistical models in predicting seedling regeneration. Tree density, stand density index, and acorn number played an important role, influencing the cork oak seedling prediction. In conclusion, the results of this research determined the importance of an AI classification modeling technique in the prediction of cork oak regeneration, providing practical references for future forest management strategy decisions.

1. Introduction

For decades, the use of artificial intelligence (AI) in various fields (from healthcare to environmental sciences) was limited by computational constraints and skepticism about its applications [1,2,3]. Nowadays, there is a huge interest in AI given its enhanced ability to process big data, identify patterns, and make predictions. Over time, the skepticism turned into the recognition of work on AI methodologies. Notably, in 2024, John Hopfield from Princeton University and Geoffrey Hinton from the University of Toronto, who pioneered tools for understanding neural networks (NN) as far back as 1982, were awarded the Nobel Prize in Physics [4,5,6,7]. At that time, AI models were constrained by the rudimentary hardware available including basic graphics support provided by cards like the Graphics Controller Adapter CGA [8]. Notwithstanding these constraints, their pioneering efforts established the groundwork for contemporary developments in AI. Consequently, the exponential enhancement in computer component capabilities and the corresponding advancements in AI technologies have facilitated the resolution of complex issues that are challenging to evaluate using non-AI models. In contrast to AI models, non-AI models demonstrate less accuracy and precise recall due to their inherent rigidity, as rule-based systems rely on predetermined rules and logic, making them inflexible in processing unstructured data [9]. Similarly, models as well as linear regression have set decision boundaries, which can pose difficulties for nonlinearly separable data [9]. Moreover, the non-AI model frequently employs static datasets, constraining its capacity to adapt to emerging patterns when these datasets are updated with new information. Consequently, within the existing scientific framework, the incorporation of AI models across diverse domains (e.g., engineering, medicine, informatics, physics, and environmental science) has markedly enhanced adaptive and resilient methodologies for addressing complex problem-solving, especially in contexts necessitating nonlinear and dynamic data processing. Recent breakthroughs in AI have shown the ability to enhance predictive accuracy and optimize decision-making processes, enabling innovative solutions to persistent scientific challenges. The trend of development in AI-related research is undergoing exponential expansion [10]. Scopus.com reports that during the period 2024–2025, there were 33,118 original scientific articles in English exclusively related to the term “artificial intelligence”, predominantly within the fields of medicine, engineering, social sciences, biochemistry, and genetics, with a lesser representation in environmental sciences and agriculture.

In particular, forestry research with AI is an emerging domain, analyzing ecological and management forest issues such as weather forecasting, forest monitoring, and wildfire detection [11,12] and has become extremely useful for database classification problems [13,14]. However, AI-based approaches in reforestation and regeneration research are still limited in employment. While the ecological and reproductive limitations of cork oak have been extensively studied using conventional statistical methods [15,16,17,18], the utilization of AI techniques, such as MLP, for predicting cork oak regeneration is still a relatively unexamined area of study. However, a focus on the functional verification of the prediction accuracy of forest regeneration modeling is still lacking.

Nowadays, the capabilities of NN in both environmental and forestry studies have been widely increased [19,20,21,22]. The NN provides a mechanism for forecasting nonlinear and categorical data, and it is intriguing to assess how NN can address challenges with enhanced accuracy and precision-recall. Furthermore, given that NN is adept at addressing multicollinearity issues among variables (input–output), this versatility applies to the analysis of ecological data, hence facilitating the prediction of seedling density based on readily obtainable data. In many forest studies with NN, the methods most used are the single hidden layer of multilayer perceptron (MLP). Already used in biology, ecology, and environmental studies, the MLP represents a robust analysis for resolving classification problems [23,24]. However, due to the complexity of the analysis, the limitations of computer calculation, and the difficulty of compression, MLP (like other type of NN) had not found a place in forestry topics until recently. Currently, MLP is being used more frequently, and due to the time lag of knowledge accumulated, further studies are needed to verify its performance in forestry disciplines. In particular, there is a significant lack of knowledge about the use on NN in modeling the natural regeneration dynamics of forest species. Selected as a case study, the authors focused on the regeneration of cork oak (Quercus suber L.) species in a Mediterranean environment, elucidating the utility of utilizing MLP to model natural regeneration during a defined period.

Cork oak is an evergreen oak that grows in the western Mediterranean Basin is and more abundant in the Iberian Peninsula [25,26], where it plays a crucial role in sustaining both environmental regulation and local economy [25,27]. Portugal has the largest cork oak cover, spanning 719.9 thousand hectares (22.3% ± 1% of mainland Portugal’s forests [28]) with a major bioeconomic impact [25,29]. The species is well-adapted to the Mediterranean climate, coping with hot and dry summers. Evolution has led the cork oak to have a high thermotolerance and ability to acclimatize to drought [30]. Cork oak’s natural regeneration consists of different reproductive stages (e.g., flowering, seed dispersal, seedling emergence, seedling survival, establishment and sapling growth). Each stage is constrained by various biotic and abiotic factors [31,32], including the vapor pressure deficit (VPD), minimum absolute temperature (Tn), canopy density, tree age, and diameter, are incorporated as factors influencing acorn development and seedling establishment. Additionally, variability across measurements and locations is introduced by factors such as predation, tree health, seed dispersal by jays, soil mineral availability, and insect infestations, significantly impacting the model predictions and reliability of ecological evaluations [17]. All of these effects may be evaluated as random effects or latent effects if the values are not available. Cork oak has unisexual reproduction with monoecious flowers. The key challenge of the cork oak reproductive system is the high degree of self-incompatibility due to the lack of synchronism between male pollen release and female flower receptiveness, which affects fertilization [33,34]. Trees begin to flower between March and June, where the wind acts mainly in pollination, and the seed dispersal of cork oak acorns between October and February is limited around individual trees [35]. An adequate seed supply starts in trees aged 15–20 years old [36].

Environmental factors, particularly low temperatures and aridity, substantially influence acorn production, leading to annual and biennial variations in yield [37]. Furthermore, acorn production varies among individual trees, and constraints in dispersion, together with post-dispersal mortality due to predation, further constrain spontaneous regeneration [35]. Then, the negative effects of climate anomalies, such as extreme heat, prolonged drought, and precipitation events (severe storms with high rainfall) lead to a general deterioration in the seedling establishment conditions and increase the susceptibility to pests and diseases, weakening the regeneration system [38,39]. In this context, the development from acorn to seedling (seedling growth) is becoming increasingly challenging for forest managers, requiring appropriate measures to ensure successful regeneration. The early prediction of the endogenous variables of oak regeneration could improve the success rate of natural or artificial plantations, although, the definitions of seedling size and survival at the beginning of germination should also be better considered.

Together with the reproductive and ecological challenges linked to cork oak regeneration, there is a deficiency in predictive understanding regarding the initial stages of this process, particularly in recognizing the factors that influence seedling survival and establishment. Currently, only a few research studies have employed AI models in cork oak forests. Studies in Portugal used data mining to predict oak habitats [40] and applied machine learning to identify pastures using remote sensing [41]. To our knowledge, there have been no studies that used neural networks or other AI models to predict how many cork oak seedlings will grow, especially when the prediction is based on forestry measurement data (e.g., using stand density indicators). While the ecological and reproductive limitations of cork oak have been extensively established using conventional statistical methods, the utilization of AI techniques, such as MLP, for predicting cork oak regeneration is still a relatively unexamined area of study. Utilizing AI to understand the limits has the potential to improve natural regeneration strategies and forest management. Specifically, the AI research on development from acorn to seedling should provide challenges for forest managers, necessitating suitable strategies to guarantee effective regeneration. Certainly, the initial predictions derived from a precise AI methodology regarding endogenous variables of oak regeneration could enhance the success rate of both natural and artificial plantations.

Our study aimed to determine which factors have the most substantial impact on the natural regeneration densities by using MLP models. The second goal was to determine the capability of MLP models to classify the cork oak seedling density within subplots with the maximum accuracy and precision-recall for a comparison with traditional quantitative approaches. We hypothesized: (i) environmental factors strongly drive the natural regeneration establishment, and (ii) MLP models perform properly, predicting natural regeneration with increased unbiasedness and accuracy.

2. Materials and Methods

2.1. Study Area and Experimental Design

The study was conducted in a permanent forest cork oak site linked to the Agenda TransForm project, which focuses on the sustainable management and the regeneration of cork oak [42]. The study site is located in Mogadouro (Bragança region), in northeastern Portugal (Lat 41.3525 N; Long −6.7617 W). The region is situated at an elevation of 570 m above sea level, is characterized by dystric and umbric leptosols, and exhibits a typical Mediterranean climate with hot, dry summers and cold, damp winters. The mean annual precipitation is 558 mm, with the rainy season spanning eight months from October to May, peaking in November and December [43].

The measurements presented in this study were carried out in 2022 and 2023. The experimental design consisted of two circular plots (A1 and A2) with an area of 500 m² (radius = 12.62 m). In each circular plot, we used two types of inventory sampling designs: (1) linear transect and (2) two-step radial cluster. (1) The linear transect design had 12 square subplots, yielding a total of 24 subplots when accounting for both plots. Each subplot encompassed an area of 1 m² (1 m × 1 m), arranged sequentially in a rectangular structure measuring 12 m in length and 1 m in breadth. The center of the linear transect was situated at the midway of each plot (Figure 1). (2) The radial cluster design consisted of five clusters of subplots, each covering a square area of 4 m². Each cluster consisted of four-square subplots measuring 1 m² (1 m × 1 m), resulting in a total of 20 subplots (4 subplots × 5 clusters) in each plot, and an overall total of 40 subplots for plots A1 and A2 combined. One cluster was located at the center of the plot, while the other four clusters were placed radially, equidistantly positioned from the center (6 m), the plot boundary, and from each other [44]. Due to the presence of two types of inventory sampling designs, we evaluated the cumulative area of overlap. This overlap was 8 m² out of 64 m² (12.5%), inclusive of all plots. We did not eliminate the overlapping observations, since recent statistical studies conducted on the same plot on the overlap issues by Fierravanti and Fonseca (submitted, [44]) indicated that the inclusion or exclusion of these measurements, representing a 12.5% design overlap and a 10% sample measurement overlap relative to the total, yielded non-statistically significant differences in the findings. The overlap area, mostly located in the plot center and encompassing half of the central cluster (refer to Figure 1), did not create statistical bias.

2.2. Regeneration and Tree Samplings

The number of acorns (Acorns) and total number of living and dead seedlings per m² (TS and DS, respectively) were recorded in two circular plots (A1 and A2). In particular, we sampled the study plots on 46 DOY (February 2022), 157 DOY (June 2022), and 263 DOY (September 2022). In 2023, there were 4 DOY (January), 181 DOY (June), 304 DOY (October exclusively for plot A2), and 312 DOY (November). The dates were chosen to cover the flowering and seed dispersal periods of the cork oak acorns, which are between February and June and between October and February, respectively [35]. The living seedlings were discriminated on the basis of stem height measured from the base to the top (H), following a threshold of H ≤ 10 cm (TS1) and H > 10 cm (TS2). The acorns (Acorns) and Seedlings were classified in three classes: Class 0 (C0), Class 1 (C1), and Class 2 (C2), where C0 represents the absence of elements (acorns or seedlings), C1 encompasses instances with a count of acorns and seedlings ranging from 1 to 4, and C2 for both the acorns and seedlings encompasses instances with counts exceeding 4. In the study, the presence of dead seedling (PDS) with stem height (H) under 50 cm was classified in two classes (YES and NO) (

Table A1. Detailed information and percentages (total and by plots) of the categorical variables in the input and output, which were used in the multilayer perceptrons (MLPs).

Layer	Variables	Classes	Details	Percentages (%)
				Total	Plot A1	Plot A2
Input	Acorn	C0	Absence of acorns	35.64	41.67	29.69
		C1	Presence between 1 and 4 acorns per m2	30.94	26.04	35.42
		C2	Presence of more than 4 acorns per m2	33.43	32.29	34.90
	PDS	No	Absence of dead seedling	90.88	89.58	89.58
		Yes	Presence of dead seedling (max 4)	9.12	10.42	10.42
	Slope	No	Flat terrain (<±5° or ~8.75%)	37.29	37.50	37.50
		Yes	Sloping terrain measured between 5° and 20° (36.4%)	62.71	62.50	62.50
Output	TS1	C0	Absence of seedling H ≤ 10 per m2	39.78	37.50	43.23
		C1	Presence between 1 and 4 seedling H ≤ 10 per m2	39.78	42.19	36.46
		C2	Presence of more than 4 seedling H ≤ 10 per m2	20.44	20.31	20.31
	TS2	C0	Absence of seedling with H > 10 per m2	37.85	20.31	56.77
		C1	Presence between 1 and 4 seedlings with H > 10 per m2	48.07	56.25	38.54
		C2	Presence of more than 4 seedlings with H > 10 per m2	14.09	23.44	4.69

). The dead seedlings during each sampling data were between 0 and 2 (98.5% of cases), and 1.5% was between 3 to 4, therefore, due to the reduced number of values up to 2, we decided to only use two classes.

To describe the stand structure, a circular plot with an area of 500 m² was established, within which all trees with a diameter equal to or greater than 7.5 cm were recorded, along with their spatial location (distance and azimuth relative to the center). Each plot was then subdivided in four quadrants (Q1, Q2, Q3, and Q4) of 125 m² each, arranged in a clockwise direction. The number of trees in each quadrant was assessed, and the density was extrapolated to a per-hectare unit, allowing for variability assessment within the plot (Figure 2).

The variable stand density index (SDI) was calculated using the expression:

$$SDI=TPH{\left({\displaystyle \frac{dg}{25}}\right)}^b$$

where TPH is the number of trees per hectare, dg is the quadratic mean diameter of the trees, and b is the allometric coefficient whose value for cork oak is defined as −1.806 [17,45]. In this application, the calculation of the index was carried out for the subset of mature trees, identified here as those with a minimum diameter of 25 cm.

We also characterized the surrounding environment, measuring the slope of each plot, classified as the absence or presence of 5° (≈8.74%) inclination, then further classified into two classes (YES and NO) (Table A1). The measurements were made at the same time of the acorn and seedling samplings.

The RND was derived through a multi-step process: (1) Exploratory factor analysis (EFA) using the maximum likelihood estimation with Varimax rotation; (2) assessment of the sampling adequacy (Kaiser–Meyer–Olkin ≥ 0.60) and sphericity (Bartlett’s Test, p < 0.05); and (3) validation of the factor structure via confirmatory factor analysis (CFA) using the CFI (>0.95), TLI (>0.95), NFI (>0.90), RMSEA (<0.08), SRMR, and χ² (p ≥ 0.05). The RND values for each plot, along with the normalized variables contributing to it, are detailed in

Table A2. Distribution of each parameter used to determine the random normal distribution (RND) obtained by the CFA. Acorns, TS1, TS2, and dead seedlings in this case were normalized continuous numeric variables. SD is the standard deviation, SE is the standard error, U 95% is the upper 95% mean and L 95% is the lower 95% mean. The values of the variables were normal in the case of RND and normalized in the case of components of RND.

Dataset	Variables		Statistics
			Mean	SD	SE	U 95%	L 95%
Total	RND		0.00	0.99	0.11	0.24	−0.24
	Components of RND	Acorns	8.00	4.90	0.61	9.22	6.78
		TS1	20.47	20.17	2.52	25.51	15.43
		TS2	20.47	20.17	2.52	25.51	15.43
		Dead seedling	1.75	2.48	0.31	2.37	1.13
Plot A1	RND		−0.01	0.93	0.16	0.32	−0.35
	Components of RND	Acorns	7.25	4.17	0.74	8.75	5.75
		TS1	20.61	19.5	3.45	27.64	13.58
		TS2	26.46	18.02	3.19	32.95	19.96
		Dead seedling	1.27	2.72	0.48	2.25	0.29
Plot A2	RND		0.01	1.04	0.18	0.39	−0.36
	Components of RND	Acorns	8.75	5.51	0.97	10.74	6.77
		TS1	20.33	21.12	3.73	27.95	12.72
		TS2	18.54	20.4	3.61	25.90	11.19
		Dead seedling	2.23	2.14	0.38	3.00	1.46

. Factor selection was based on eigenvalues, with Factor 1 (eigenvalue = 2.52) explaining 62.01% of the variance. Factor 2 (eigenvalue = 0.8) was excluded. Variables retained from Factor 1 for CFA included TS1 (0.94), TS2 (0.57), Acorns (0.67), and dead seedlings (0.66) (Table A2).

2.3. Multilayer Perceptron Analysis

The multilayer perceptrons (MLPs) used in this study were constructed with a block of five neurons each, employing the hyperbolic tangent (TanH), linear (LIN) and Gaussian (GAUS) activation functions across two consecutives hidden layers (H1 and H2), from input to output (Figure 3).

Consistently, the number of neurons remained uniform across both layers. A total of 10 MLPs models were tested (Figure 4), ranging from a minimum of 30 neurons (e.g., MLP-5 = (5 TanH + 5 LIN + 5 GAUS neurons) × 2 hidden layers) and a maximum of 300 neurons (e.g., MLP-50 = (50 TanH + 50 LIN + 50 GAUS) × 2). The use of these three activation functions (TanH, LIN, and GAUS), dispersed across hidden layers with a comparable number of neurons, improved the model’s capacity to generalize and forecast regeneration patterns. The TanH function translates inputs to a range of −1 to 1, encapsulating intricate relationships. The LIN preserves linearity, which is advantageous when input–output interactions are proportional. GAUS, characterized by its bell-shaped curve, represents localized responses and gradual fluctuations. This balanced amalgamation utilizes nonlinearity (TanH and GAUS) and linearity (LIN), guaranteeing consistency and stability while averting bias toward any singular activation function. In the context of classification, the integrated use of TanH, LIN, and GAUS enables the model to adjust to various data structures and enhances generalization [46,47]. To assess the predictive performance of the model, each MLP model underwent initial validation using a portion of the dataset generated through the k-fold method [48,49,50]. In our study, we employed a k-fold value of 6, resulting in a cross-validation dataset comprising 16.6% (83/17 split) of the total data, as it balanced the accuracy between k = 5, which performed poorly on validation and training (GR²), and k = 7, which affected the prediction quality. The MLPs’ tested inputs were slope (surrounding environment, categorical variable), number of trees per hectare (TPH), stand density index (SDI) (stand tree variables), DOY (time, categorical variable), PDS, and number of acorns classified (regeneration, categorical variable), and RND (random numerical variable), while the output of the models was the seedling categorized by height (H) in TS1 (H ≤10) and TS2 (H > 10) (regeneration categorical variable) (Figure 4).

2.4. Performance Analysis of MLP Models

The performance of the MLP models was assessed systematically in five steps (Figure 2): (1) measuring the accuracy of the MLPs; (2) measuring the precision; (3) measuring the recall (sensitivity); (4) identifying whether there were inputs that were problematic in the prediction of output; and (5) comparisons with other AI and non-AI statistical analyses. In general, one method to obtain an initial understanding is the confusion matrix, which was used to gain insights into the MLPs’ accuracy, precision, and recall. The confusion matrix was employed to assess the misclassification in both the training and validation predictions [51,52]. A comprehensive metric analysis was conducted to thoroughly examine these indicators of model prediction accuracy. Eight metric parameters were utilized as indicators to evaluate the model’s fit regarding the accuracy, precision, and recall of the MLP models for both the validation and training datasets (please refer to

Table 1. List of metric parameters used to evaluate the accuracy and precision of the MLP model, along with their interpretations and relevant references (Ref.).

Metric Parameter	Abbr.	Definition	Interpretation	Ref.
Generalized R2	GR2	Measures model performance; higher values close to 1 indicate better performance, while negative values imply worse performance compared with a model predicting the data’s average.	Higher GR2 values indicate superior model fit, while negative values reflect a poor performance.	[53]
Entropy R2	ER2	A variant of GR2 based on entropy theory; higher values close to 1 represent better performance.	Similar to GR2, higher values signify improved accuracy and model fit.
Negative log-likelihood	NLL	Evaluates the likelihood of the model predictions given the data; lower NLL values denote a better fit.	Smaller NLL values reflect a superior model performance.	[54]
Root average squared error	RASE	Calculates the square root of the mean squared differences between the predictions and actual data values.	Smaller RASE values indicate a better predictive performance.
Mean absolute deviation	MAD	Computes the average absolute difference between the predictions and actual data values.	Lower MAD values suggest a higher model accuracy.	[55,56]
Misclassification rate	MR	Proportion of misclassifications to total classifications; ranges from 0 (perfect classification) to 1 (all data misclassified).	Smaller MR values denote better a classification accuracy.
Receiver operating characteristic (ROC) area under curve (AUC)	ROCAUC	The area under the receiver operating characteristic curve; it evaluates a binary classification model’s performance by plotting the true positive rate against the false positive rate at various thresholds.	AUC ranges from 0.5 (random classification) to 1 (perfect classification); higher values indicate better discrimination between classes.	[51,57]
Precision-recall (PR) area under curve	PRAUC	Evaluates the precision and recall for imbalanced datasets, focusing on positive class performance; ranges from 0 to 1.	A PRAUC value of 1 indicates perfect precision and recall, highlighting strong performance, particularly in imbalanced data scenarios.	[51,52,58]

for further information).

2.5. Comparison Between MLP Models and Other Prediction Statistical Models

The effectiveness of MLP in terms of modeling and performance was assessed by comparing the same accuracy, precision, and recall indicators used in the comparisons among the MLPs. For the comparison, we chose to compare the MLP with the most widely used AI and non-AI methodology as well as studies in various sectors of forestry ecology. Therefore, as AI modeling methods, we chose to use bootstrap forest (BT), decision trees (DT), k-nearest neighbor (kNN), Naive Bayes (NB), support vector machine (SVM), and neural network boosted (NNB). Notably, the NNB models consisted of single-layer architectures with three neurons utilizing the TanH activation function. Specific parameters for boosting were defined including a maximum of 100 interactions and a learning rate to 0.1. In the case of non-AI models, we chose the nominal logistic (NL) and generalized regressions with Lasso (GRL), Elastic Net (GRE), and Bridge (GRB) estimation methods. To identify a more suitable traditional model for comparison with the MLPs, the NL and generalized regressions models were further evaluated against each other using Akaike information criterion correct (AICc) and Bayesian information criterion (BIC) indicators. Moreover, the fit of the NLs and their source of variation was further evaluated in detail using the square tests (χ²). Similar to the MLPs, a k-fold of 6 was employed for these statistical analyses.

A detailed analysis based on the G² (the likelihood ratio chi-square statistic) was carried out for the BF. The G² is a statistic used to evaluate how well a split (or division) works within a decision tree. The G² measures the difference between the expected and the observed data distribution after the split. A higher G² value indicates a better separation between the target classes. When constructing BF trees, the G² is calculated for each possible split, and the one that maximizes this statistic is selected. The best MLP, AI, and traditional statistical (non-AI) prediction models were also visually compared with a mosaic plot to check the accuracy.

2.6. Assessing the Effects of Inputs on Outputs

An independent resampled input (IRI) technique was used to identify the role of the surrounding environment, stand characteristics, regeneration variables, time, and random inputs on the prediction of the seedling density output after the best MLP was used. This methodology permitted us to assess the effects of different inputs while maintaining their independence through resampling methods on the TS1 and TS2 [59]. From the IRI, we considered the total effect, which is the comprehensive impact of a predictor variable (e.g., Acorns) on the input (TS1 and TS2) including all direct and indirect pathways and the main effect, which is the direct impact of a predictor variable on the response, without considering the interactions or the effects of other predictors [59]. All analyses performed in this research were conducted using JMP^®, Version 18 Pro, Student Edition (SAS Institute Inc., Cary, NC, USA, 1989–2024).

3. Results

3.1. Performance Analysis of the MLP Models

Our study showed that the MLP-40 and MLP-50 multilayer perceptron models best showed regeneration, as their GR² and NLL values were similar. The advanced GR² in MLP-40 showcased the model’s capacity to clarify data discrepancies. MLP-40 demonstrated more accuracy than MLP-50 (which has more neurons) in data classification (Figure 5). The GR² of MLP-40 varied between 0.47 and 0.57, depending on the training or validation dataset. GR² had a variation of ±0.06 between the training and validation datasets (

Table A3. Performance of MLPs in predicting the seedling height (H) categorized as H ≤ 10 (TS1) and H >10 (TS2). Metric parameters GR² and ER² were generalized R-squared and the entropy and NLL was the log-likelihood. RASE represents the root mean square error, while MAD and MR denote the mean absolute deviation and the misclassification rate, respectively. (*) represents the least effective MLP model, while (**) denotes the most accurate MLP model. ‘NA’ indicates that the result is not available.

Response	Model	Training					Validation
		GR2	ER2	RASE	MAD	MR	GR2	ER2	RASE	MAD	MR
TS1	MLP-40 **	0.67	0.42	0.45	0.36	0.27	0.74	0.51	0.42	0.34	0.21
	MLP-5 *	0.45	0.24	0.54	0.49	0.37	0.4	0.18	0.56	0.51	0.41
	BF	0.63	0.38	0.48	0.45	0.26	0.06	0.03	0.68	0.56	0.45
	SVM	0.42	0.22	0.54	0.49	0.37	0.24	0.12	0.56	0.52	0.38
	NNB	0.36	0.18	0.56	0.52	0.4	0.27	0.13	0.56	0.52	0.4
	DT	0.35	0.17	0.56	0.53	0.4	0.25	0.12	0.56	0.52	0.36
	NB	0.32	0.16	0.57	0.53	0.4	0.16	0.07	0.58	0.54	0.38
	kNN	NA	0	NA	NA	0.51	NA	−0.21	NA	NA	0.48
	NL	0.34	0.17	0.57	0.53	0.38	0.21	0.1	0.57	0.53	0.42
	GRL&GRN	0.3	0.15	0.58	0.55	0.43	0.21	0.1	0.58	0.58	0.36
	GRR	0.34	0.17	0.57	0.53	0.39	0.21	0.1	0.57	0.53	0.41
TS2	MLP-40 **	0.58	0.35	0.47	0.4	0.28	0.63	0.43	0.47	0.4	0.28
	MLP-5 *	0.35	0.35	0.59	0.53	0.41	0.28	0.13	0.58	0.55	0.45
	BF	0.57	0.33	0.5	0.48	0.28	0.12	0.05	0.6	0.57	0.48
	SVM	0.34	0.17	0.56	0.52	0.39	0.17	0.07	0.6	0.56	0.47
	NNB	0.28	0.13	0.58	0.56	0.44	0.21	0.1	0.59	0.57	0.43
	DT	0.16	0.07	0.6	0.59	0.49	0.17	0.07	0.6	0.58	0.45
	NB	0.07	0.03	0.58	0.55	0.52	−0.02	−0.01	0.58	0.55	0.5
	kNN	NA	−0.05	NA	NA	0.54	NA	0.05	NA	NA	0.45
	NL	0.28	0.13	0.58	0.55	0.44	0.18	0.08	0.59	0.57	0.43
	GRL&GRN	0.27	0.13	0.58	0.56	0.45	0.2	0.09	0.59	0.57	0.44
	GRR	0.26	0.12	0.59	0.57	0.44	0.19	0.08	0.6	0.57	0.42

). The model accurately predicted the growth of short, high-density seedlings in TS1 better than that of tall, low-density seedlings in TS2. We achieved this by examining the seedling density and various other metrics. There were the same number of observations in both the training dataset and validation dataset. However, the validation dataset had 80% less mean negative log-likelihood (NLL) than the training dataset. The NLL differences in the training datasets for TS1 and TS2 were about 5%–8% on average. Predictions for short seedlings were lower, which showed that TS1 had better MLP-40 accuracy for seedling density classifications per square meter (C1, C2, and C3). The MLP-40 demonstrated accuracy, as evidenced by the training and validation confusion matrix (

Table 2. Confusion ratio of the MLP-40 model. Accuracy of the MLP-40 model in classifying the three seedling classes (C1, C2, and C3) based on the output (TS1 and TS2); actual values were obtained during the field sampling and predicted by MLP-40. The numbers written in bold are related to the correct prediction in terms of the actual-predicted ratio.

	Output	Actual	Predicted Rate
			Training			Validation
			C0	C1	C2	C0	C1	C2
MLP-40 model	TS1	C0	0.82	0.13	0.04	0.94	0.03	0.03
		C1	0.17	0.77	0.07	0.16	0.79	0.05
		C2	0.14	0.38	0.48	0.17	0.33	0.50
	TS2	C0	0.75	0.20	0.05	0.67	0.29	0.04
		C1	0.15	0.78	0.07	0.07	0.86	0.07
		C2	0.11	0.38	0.51	0.24	0.29	0.48

). In addition to the GR² and NLL, the results showed that MLP-40 was better than all other MLP models in every way (Figure 5). These results included model accuracy, prediction error, variance explanation, misclassification, and deviation. As shown by ER², MLP-40 was better at explaining the variance in the data. It also produced lower RASE and MAD values, which means that it was more accurate at making predictions and had fewer errors. Furthermore, MLP-40 had a reduced MR, signifying enhanced classification efficacy.

The MLP models with the highest and lowest neuron counts (MLP-50 and MLP-5) exhibited suboptimal performance. The MLP-50 model’s increased NLL fit with more neurons suggests difficulties in data modeling. The performance of the MLP-50 network might indicate overfitting due to the increased number of neurons, but this was not found, as the GR² of the training set was equal to or slightly lower than that of the validation set, indicating that the model generalized well, rather than memorizing the training data. Despite MLP-50 possessing 60 more neurons compared with MLP-40, the GR² and NLL fit–accuracy trade-off indicates an imbalance. Analogous to the GR²–NNL connections, MLP-5 forecast seedling quantities with the least precision and fit. MLP-5 was less complex to compute than MLP-50 due to having 10 times fewer neurons. The GR²–NLL relationship indicates that MLP-40 optimally balanced model accuracy and efficiency (Figure 6).

3.2. Comparison of the Best MLPs with Other AI and Non-AI Models

The assessment of the model’s effectiveness in distinguishing the seedling counts at different choice thresholds validated the compromise between sensitivity and specificity for the area under the ROC curve (ROC_AUC). MLP-5 was recognized as the least effective model, and MLP-40 and MLP-50 were designated as the most accurate AI models in this study. Moreover, MLP-40 exhibited an enhanced performance on the training datasets, attaining the best overall accuracy, which was attributed to the increased ROC_AUC values (Figure 7), whereas MLP-50 displayed inferior ROC_AUC values, indicating a 10% difference in the ROC curve and true positive sensitivity relative to the less complicated MLP-40.

The ROC_AUC value aligns with the other metrics used for comparing AI and non-AI models, reinforcing their applicability in resource-efficient scenarios (Figure 8).

Furthermore, MLP-40 showed significant effectiveness versus the most challenging AIs such as NB and kNN. The Naive Bayes and k-nearest neighbor algorithms exhibited suboptimal performance due to elevated entropy (ER²). Within the metric parameters, the precision of the AI, bootstrap forest (BF), and the other neural network category, NNB, was comparable to that of the MLPs. Nonetheless, the BF model often generated erroneous positive predictions, particularly with the prediction of taller seedlings (TS2) in count class C2 or the representative class, which denotes more than four seedlings per square meter. Misunderstandings on the false positive predictions of BF persisted across the validation dataset. Nonetheless, it was established that MLP-40 surpassed BF, while confusion was observed in MLP-40 (Figure 9).

3.2.1. Comparison Between MLP Models and Other Dedicated AI Models

As above-mentioned, the MLP models demonstrated a superior performance across various metrics, and their high accuracy and precision-recall were well-documented by the high values of the AUC. This was followed by BF, the other AI, which was competitive when focusing on training datasets, and certain models, such as NNB and SVM, showed moderate predictive performance. However, there were more negative differences in the classification of C2 seedlings, particularly between the MLPs and the NB, or DT or kNN models. The BF model also achieved a robust level of precision, as shown by PR_AUC values in the training phase (up to 0.88), although it slightly declined during validation. The study of other several AI models showed that improvements to the neural network (e.g., NNB models) did not make them as accurate as the more complex MLP models. The GR² for the NNB remained largely unchanged at 0.27 during training and 0.20 during validation. While NNB is a faster one-layer neural network model for prediction than MLP-40 and is somewhat good at boosting, it struggled with classification accuracy, especially in the C2 category (seedling number above four), which is a problem that BF also had (

Table 3. Confusion ratio of the NNB model. Accuracy of the NNB model in classifying the three seedling classes (C1, C2, and C3) based on the output (TS1 and TS2); actual values were obtained during the field sampling and predicted by NNB. The numbers written in bold are related to the correct prediction in terms of the actual-predicted ratio.

	Output	Actual	Predicted Rate
			Training			Validation
			C0	C1	C2	C0	C1	C2
NNB model	TS1	C0	0.80	0.20	0.01	0.82	0.16	0.02
		C1	0.28	0.69	0.03	0.20	0.78	0.03
		C2	0.18	0.71	0.11	0.25	0.70	0.05
	TS2	C0	0.60	0.40	0.00	0.60	0.41	0.00
		C1	0.26	0.75	0.00	0.34	0.66	0.00
		C2	0.09	0.91	0.00	0.11	0.90	0.00

). The MLP-40 and MLP-50 models did well on both the training and validation datasets for TS1 and TS2. This was shown by their ROC_AUC scores, which measured how accurate they were. Neglecting the MLP models, the BF model was the best AI model, obtaining the same ROC_AUC value of 0.92 as MLP-50 when using the training dataset. Nonetheless, the NNB and support vector machine (SVM) models exhibited modest ROC_AUC values of 0.82 and exhibited modest PR_AUC values of 0.69 and 0.70, respectively, in terms of precision-recall. The study ultimately confirmed that the C2 group faced the greatest challenge in predicting seedling outcomes. In comparison to BF’s accuracy and precision, the NB, DT, and KNN models performed poorly in categorizing the C2 category across all seedling types, particularly in terms of precision and recall (Figure 9).

3.2.2. Comparison Between MLP Models and Non-AI Models

The non-AI models did not surpass the MLP-40 in accuracy and precision-recall during the predictive tasks. The ROC_AUC and PR_AUC values, which were inferior to those of MLP-40, suggest that the non-AI model was less effective in categorizing seedlings. The least powerful MLP-5 model outperformed the non-AI statistical methods, demonstrating superior capability in managing imbalanced class distributions. The goodness-of-fit uncertainty measurements (entropy) in classification via GR² and ER² indicated values that were 22%–33% lower in the non-AI models compared with MLP-40, hence affirming the enhanced predictive capability of AI models (Table A3). The confusion matrices indicated that the nominal logistic (NL) models consistently misclassified C2 as C1 across all seedling dimension categories (Figure 8). The MLP models, in contrast, had superior accuracy in classifying the items properly.

3.3. Main Effect of the Input Factors

The independent resampling input (IRI) of the MLP-40 model indicated that the stand variables associated with tree density and the stand density index (SDI) for mature trees over 25 cm in diameter had the most pronounced influence on the seedling density prediction (main effect 25%, total effect 51%) (Figure 10). The second variable that defined the response of the number of trees per hectare (TPH) (main effect 16%, total effect 40%) (Figure 10).

The quantity of acorns in terms of absence, scarce, or moderate presence was the third most significant factor for the prediction of seedlings. The total effects of time and slope were also considerable, ranging from 22% to 17%, respectively, for the seedling density, while them main effects were 2% and 3%, respectively. Furthermore, among the variables influencing the seedling density over time, the presence of perished seedlings demonstrated a relatively minor effect (main effect 12%, total effect 2%). The RND may accommodate random effects absent from the major stated variables. In comparison to the other non-AI models, the NL model demonstrated the optimal statistical fit for both shorter and taller seedlings (TS1 and TS2), as indicated by Δ (AICc) and Δ (BIC) values of zero; hence it was classified as NL (

Table 4. Akaike information criterion correct (AICc) and Bayesian information criterion (BIC) indicators concerning the nominal logistic (NL) and generalized regression models (Lasso, Ridge, Elastic Net). The Δ (AICc) and Δ (BIC) were the difference between the AICc and BIC of the best model and other models for each response (TS1 and TS2). The values of 0.00 in Δ (AICc) and Δ (BIC) represent the best models.

Response	Models	AICc	Δ (AICc)	BIC	Δ (BIC)
TS1	NL	790.36	0	869.06	0
	GRL	808.91	−18.55	880.07	−11.01
	GRE	811.13	−20.77	886.14	−17.08
	GRR	813.72	−23.36	892.58	−23.52
TS2	NL	800.02	0	878.88	−38.52
	GRL	806.14	−6.12	861.77	−21.41
	GRE	820.26	−20.24	840.36	0
	GRR	814.58	−14.56	893.44	−53.08

). Furthermore, the overall model test of NL demonstrated a significant χ²_(Wald) with a p < 0.0001 for both models with seedlings in TS1 and TS2 as the responses, indicating that the predictors of TS1 and TS2 are associated with the actual responses of TS1 and TS2. Conversely, the lack of fit test results indicated a χ²_(Wald) that was not significant (p > 0.05) (

Table A4. Presentation of the nominal logistic (NL) “whole model test” and “lack of fit” results. The χ² is the chi-square of the log-likelihood with its p-value (p).

Response	Metric Indicators	Whole Model Test	Lack of Fit
TS1	χ2	121.85	768.43
	p	<0.001	0.91
TS2	χ2	110.39	769.32
	p	<0.001	0.91

), suggesting that the model adequately fit the data.

The comprehensive model test validated that NL may elucidate phenomena (χ²_Wald, p < 0.0001), and the findings offer partial endorsement for MLP-40’s IRI. The NL demonstrated that Acorns were the primary variable for TS1 in all models (χ²_Wald: 73.03–78.37, p < 0.0001). SDI and TPH significantly influenced TS2, particularly in the generalized regression Lasso (GRL), Elastic Net (GRE), and Ridge (GRR) models. SDI had more significance than Acorns (χ²_Wald: 57.80–57.95, p < 0.0001) (

Table 5. Effect tests of nominal logistic (NL), generalized regressions (GRs), Lasso (GRL), Elastic Net (GRE), and Ridge (GRR). The χ²_(Wald) is Wald’s chi-squared significance test with its p-value: p ≥ 0.05 (ns), p < 0.05 (*); p < 0.01 (**), p < 0.0001 (***).

Response	Source of Variation	χ2(Wald)
		NL	GRL	GRE	GRR
TS1	SDI	0.00 ns	28.62 ***	21.51 ***	9.35 **
	TPH	0.00 ns	77.45 ***	71.01 ***	57.58 ***
	Slope	6.41 *	6.39 **	6.39 *	6.35 *
	Acorns	78.37 **	73.2 ***	73.19 ***	73.03 ***
	PDS	5.37 ns	5.08 ns	5.08 ns	5.09 ns
	DOY	0.84 ns	0.82 ns	0.82 ns	0.85 ns
	RND	0.22 ns	0.21 ns	0.21 ns	0.22 ns
TS2	SDI	0.00 ns	57.95 ***	57.80 ***	12.59 ns
	TPH	0.00 ns	6.78 *	7.03 *	57.50 ***
	Slope	2.33 ns	1.08 ns	1.16 ns	2.24 ns
	Acorns	24.27 ***	19.48 ***	19.66 ***	25.57 ***
	PDS	2.99 ns	1.84 ns	1.95 ns	3.33 ns
	DOY	8.49 **	6.03 **	6.17 *	8.15 *
	RND	1.65 ns	0.31 ns	0.39 ns	1.66 ns

).

The NL and regression models had similar GR² and ER² values, differing by around 4%; however, they remained inferior to MLP-40. The BF results, closely resembling those of MLP-40, indicated that, similar to MLP-40, the presence of acorns significantly influenced seedling selection. However, in contrast to MLP-40, timing (or, in this instance, the DOY) significantly influenced the model predictions. The acorns significantly influenced tiny seedlings (TS1), yielding a G² score of 71.94. Nonetheless, they still influenced bigger seedlings (TS2), yielding a G² score 53% less than TS1. Moreover, the temporal variable DOY strongly affected all groups of seedlings, exhibiting similar G² values in TS1 and TS2 (

Table 6. Bootstrap forest (BF) model splits, G² is the likelihood ratio chi-square statistic, and portion results.

BF Model	Output	Input	Splits	G2	Portion
	TS1	SDI	51	12.77	0.05
		TPH	61	9.73	0.04
		Slope	94	17.99	0.06
		Acorns	84	71.94	0.26
		PDS	40	8.41	0.03
		DOY	349	53.13	0.19
		RND	513	103.39	0.37
	TS2	SDI	8	19.62	0.07
		TPH	11	25.21	0.09
		Slope	141	20.75	0.08
		Acorns	132	33.46	0.13
		PDS	43	8.94	0.03
		DOY	302	53.60	0.2
		RND	493	104.51	0.39

).

4. Discussion

The results of the study confirm both hypotheses, showing that the environmental factors effectively and strongly drive the establishment of natural regeneration over time. In detail, the findings corroborate the notion that stand characteristics associated with the stand density index (SDI) significantly affect the seedling density, alongside the critical impact of acorn quantity and seasonal variations. Moreover, the findings prove that the MLP outperformed earlier prediction models, including both AI and non-AI systems, in predicting natural regeneration with enhanced accuracy and precision-recall.

4.1. The Relative Importance of Main Explicative Factors on Seedling Regeneration

The study indicates that the MLP-40 IRI data demonstrated that SDI and TPH are significant variables influencing the seedling density of cork oak. These forests, marked by recurrent drought conditions, experience intensified competition for water, light, and nutrients, negatively impacting forest health and resilience. The MLP-40 model showed that the SDI and tree density of the stand had a large influence on the prediction of the classes of the number of seedlings. Even without modeling, differences in the SDI and tree density were evident across the stands. The model effectively incorporated and predicted these distinctions. Anticipating differences in the SDI and tree density are very important as they impact germination, and as a whole process, regeneration from acorn to seedling. These regeneration factors influence the availability of essential resources for growth and the survival of seedlings such as light, water, and nutrients [17,60]

The SDI, as a vital dendrometric parameter, is an essential metric for forest managers, allowing them to evaluate when a stand nears its biological carrying capacity. In this setting, management actions like thinning are essential for reducing competition, fostering tree development, and improving cork quality.

Our findings corroborate previous studies, indicating that sustaining ideal TPH or SDI values benefits both the environment and forestry [17,45]. Particularly for TPH in the previous two decades, studies indicate that TPH levels between 50% and 75% are recommended to promote natural regeneration [17]. These values ensure that sufficient light penetrates to facilitate the growth of understory vegetation and the regeneration of oaks while simultaneously reducing the fuel loads and enhancing the resilience of the stand. By adding the SDI or TPH to predictive models, forest managers can better predict and deal with the problems caused by environmental stresses, which will make the cork oak ecosystems in northern Portugal healthier and more resilient. Indeed, recent studies show that high tree density can intensify competition for resources, potentially impeding the growth and survival of seedlings [17,61,62]. Conversely, a lower tree density may afford the seedlings with ample resources, promoting their growth and survival [63]. Recent studies have highlighted the importance of these factors in oak regeneration. However, we contend that this axiom may not hold true in regions with Mediterranean climates, where rainfall and transpiration levels differ [64]. Instead, in warm areas within cork oak habitats, the density of trees and the SDI in a cork oak population can influence microclimates, resource availability, and competition, all of which are vital for seedling survival and growth [65]. An increase in tree density per hectare could promote regeneration in cork oak forests in Portugal [66,67], and higher tree densities could create a more favorable microclimate for seedling establishment and growth [68,69]. This is due to several factors; for instance, higher tree densities can provide more shade, thereby reducing the soil temperature and evaporation, which in turn conserves soil moisture [70]. One of the most significant factors contributing to this favorable microclimate is humidity.

The transpiration from a larger number of trees can increase the local humidity, which can be beneficial for seedling growth [71,72]. Increased tree density can enhance plant nutrients in soil, primarily because more trees result in greater leaf litter accumulation. As this litter decomposes, it enriches the soil with organic matter and nutrients [15,73]. Indeed, while higher tree densities can facilitate regeneration, it is essential to recognize that there is a threshold beyond which competition for resources (such as light, water, and nutrients) can become detrimental to seedling growth [74,75]. Therefore, finding the optimal tree density is crucial for promoting successful regeneration [76,77]. A model considering the impact of density on natural regeneration dynamics offers valuable insights for decision-making support in forest management.

In terms of seedling survival, several factors can contribute to acorn and seedling survival in cork oak woodlands. An interesting environmental factor identified was topography [78,79], specifically the slope of the terrain where acorns initiate the germination process, ultimately giving rise to seedlings. The MLP-40 model confirmed that the slope of the terrain plays a role in the regeneration of oaks, especially during the transition from acorn to seedling. The terrain slope can impact the water runoff, soil erosion, and solar radiation levels, all of which are pivotal factors affecting germination and growth [80]. Water is essential for acorn germination and seedling growth. However, steep slopes can exacerbate water runoff, potentially drying out the soil and rendering it unsuitable for germination. Conversely, gentler slopes have a greater capacity to retain water, fostering successful germination and growth. Soil erosion poses another challenge on sloping terrain [70]. Erosion can displace acorns and young seedlings or deplete essential nutrients from the soil [78]. Therefore, ensuring soil stability on the slopes is decisive for the successful establishment of cork oak seedlings. Finally, the orientation of the slope can influence solar radiation, which is a key factor for photosynthesis. For example, north-facing slopes in the Northern Hemisphere receive less direct sunlight than south-facing slopes, potentially resulting in cooler and wetter conditions [81,82].

4.2. The Ability of MLP to Prediction of Seedling Regeneration Compared with Other Models

The research identified superior performance in some MLPs (e.g., MLP-40 and MLP-50) for the prediction and categorization of seedlings. MLP-40, with a total of 240 neurons, had the maximum accuracy and carried out an exceptional fit to the data, as seen by the metric indicators. This outcome corroborates extensive research across several fields indicating that the utilization of MLP yields significantly high accuracy, precision, and recall [83,84,85,86,87,88]. Furthermore, our use of various activation functions, evenly distributed throughout the two hidden layers, demonstrated an enhancement in the capacity to discern complicated, linear, and nonlinear connections, along with category categorization of the data related to regeneration. The MLP results affirm that a two-hidden-layer configuration is accurate in predicting and classifying regeneration concerning seedlings, surpassing boosted neural networks and other artificial intelligence methods [89,90]. In cork oak regeneration situations, the integration of TanH, LIN, and GAUS activation functions demonstrated efficacy in predicting outputs, even in the presence of negative values, gradient-based optimization, and nonlinear input values [91,92,93]. Moreover, the study validates the significant efficacy of MLPs, including MLP-40, in comparison to conventional quantitative methods. Our work confirms the premise that MLP models effectively forecast natural regeneration, demonstrating that MLPs can model cork oak regeneration with a high degree of impartiality and accuracy, surpassing both AI and non-AI models.

In particular, MLPs have shown more flexibility and more accuracy and precision-recall than models that can only capture linear relationships, such as the nominal logistic, which is a statistical model classically used for binary classification problems. Non-AI models are simple, fast, and provide good performance when the relationship between the input and output is linear or nearly linear. The most closely related method to MLP is NNB. However, our findings revealed that NNB yields inferior metric parametric results, translating in lower accuracy. Despite NNB’s capability of forming large neural network models through boosting, with a single hidden layer of neurons, it exhibited lower accuracy compared with MLP in our specific case. The confusion matrix underscores the high classification performance of MLP-40 and MLP-50, aligning with findings from environmental research that validate their utility in assessing outcomes with high precision. In addition, the MLP results outperformed other ML statistical analysis. Both BF and SVM demonstrated weaker performance compared with MLP. BF exhibited robustness against overfitting and could handle complex nonlinear relationships [91,94,95]. Similar to MLP, it can be computationally intensive and less interpretable than simpler models. However, unlike MLP with a high number of neurons, it had less accuracy in the validation.

The MLP-40 also surpassed SVM in accuracy. SVM is a type of machine learning model that can perform both linear and nonlinear classifications as well as MLP, and it is known for its robustness and ability to handle high-dimensional data (see [96,97]). However, in our case, the SVM was less accurate than MLP.

We found that SVM mistook the prediction of classes during the training and validation. We speculate that this mistake of typology was due to the fact that SVM can be sensitive to the choice of kernel and hyperparameters and may not perform well with many overlapping classes [98]. This was the case in our instance, where SVM misclassified the number of seedlings in C2.

5. Conclusions

The research was conducted over two years in a mature Mediterranean cork oak forest in Portugal, focusing on natural regeneration and using MLP-models. Our approach highlighted the key role of stand tree density in the regeneration success of cork oak. The slope was also identified as a vital factor for seedling growth. In our study, the MLP models provided elevated accuracy and precision in recall, with a discrete number of neurons and combination of activation functions. MLP models surpassed non-AI models in recognizing the significance of acorn quantity for seedling density. The MLP-40 model demonstrated strong performance in predicting the seedling density in pure cork oak stands in northern Portugal, highlighting the stand density index (SDI) and trees per hectare (TPH) as key predictors. Finally, this study offers practical guidance for forthcoming decisions regarding forest management strategies. The determination of the optimal density values should integrate insights from this study alongside recommendations from previous research [17,45] to support the more effective management of cork oak natural regeneration.

The results emphasize the importance of managing the mature tree density in pure cork oak stands to enhance the establishment and recruitment of natural seedlings. Thinning practices should be designed to optimize tree spacing while preserving ecosystem functionality and cork productivity. Further research is recommended to validate these management strategies across different environmental conditions and investigate additional ecological factors that may influence seedling dynamics and stand structure. By integrating AI-driven models into management practices, forest practitioners can better anticipate the outcomes of different interventions, ultimately promoting sustainable cork oak regeneration in Mediterranean ecosystems.