Evaluating Machine Learning Algorithms for Modeling Forest Characteristics in Romania Using Remote Sensing Data

1. Introduction

Forests are vital ecosystems, offering a diverse array of provisioning, regulating, cultural, and supporting services [1,2,3], thus playing a critical role in sustaining human well-being. Accurate and updated information on measurable forest attributes is essential for an effective forest management [4,5]. Typically, attributes such as the standing stock (Vol), basal area (BA), and aboveground biomass (AGB), are used to understand and differentiate forest structure and productivity [6,7]. For instance, estimates of standing stocks are essential for assessing timber quantity, in guiding sustainable harvesting practices, and ensuring the long-term viability of forest resources. Additionally, BA offers insights into stand density and overall forest health [6]. By understanding the spatial distribution and density of trees within a forest stand, managers can better plan, mitigate risks of disease and fire, and maintain ecosystem resilience [8]. In turn, AGB estimates are important in understanding the forests’ potential to store carbon, which is important to shape policies for decarbonization and to mitigate climate changes.

Typically, all these parameters are estimated based on manual ground measurements, standing for the conventional approach to the problem, which often involves resource-intensive field surveys, expertise, and working in adverse conditions [9] . Furthermore, conventional methods may not be scalable or cost-effective for large-scale biomass mapping and estimation efforts, particularly in remote or densely forested areas [10,11,12]. According to the methodology outlined by the Intergovernmental Panel on Climate Change (IPCC), as well as by several local guidelines and textbooks, the AGB of trees can be estimated using allometric equations parameterized with diameter at the breast height (DBH) and tree height (H) as primary variables [13]. These equations are designed to provide accurate biomass estimates by leveraging the correlation between tree size and biomass stored in it [13,14].

Despite the straightforward and reliable nature of DBH and H measurements across various forest types and regions, these persist as resource intensive and often lack scalability [15,16]. Integrating remote sensing technologies such as satellite imagery and LiDAR (Light Detection and Ranging) are promising in removing these limitations, making forest attribute estimation more efficient and cost-effective [17]. While these methods allow for broader coverage and can capture data in otherwise inaccessible regions, they require accurate ground truth data for validation and hold a limited ability to characterize the understory vegetation [17].

Regression algorithms have long been used as powerful tools for extrapolating field data and generating detailed maps of forest attributes [18,19]. These algorithms use statistical techniques to model the relationships between various forest attributes and environmental variables [20]. Integrating them with remotely sensed data offers a promising approach to improving forest management practices [11,21]. Forest managers, for instance, can gain valuable insights into various aspects of forest dynamics, including vegetation health and biodiversity distribution [22]. This combination enables a more accurate and efficient monitoring and assessment of forests at different scales, supporting decision-making in forest conservation, sustainable resource management, and the prevention of environmental threats [23,24].

Although various machine learning algorithms have been used in estimating forest attributes, the choice between parametric and non-parametric algorithms remains challenging [25,26]. Parametric algorithms, such as linear regression, rely on several assumptions about the data distribution and hold fixed parameters [27]. On the other hand, non-parametric algorithms, including Random Forest (RM) and Support Vector Machines (SVM), are more flexible because they can capture complex relationships without strict statistical assumptions [28]. Understanding the strengths and weaknesses of algorithms from this class is essential in selecting the optimal one for specific forest attribute estimation tasks.

Among non-parametric algorithms, RF, GBTA (Gradient Boosting Tree Algorithm) and CART (Classification and Regression Trees) have demonstrated their effectiveness in accurately estimating various forest characteristics such as the above ground biomass [29,30,31], forest canopy height [32] and standing stock [33,34]. The RF algorithm minimizes prediction variance in supervised machine learning by combining multiple decision trees through a process called bagging [35]. Initially, each tree is trained on random subsets of data; during validation, the algorithm averages the predictions to come at a balanced prediction outcome [35,36]. Its effectiveness lies in handling intricate decision boundaries for both categorical and continuous variables, managing noisy and high-dimensional datasets, and minimizing overfitting [37]. The algorithm incorporates randomization techniques to reduce tree correlation, which enhances the model accuracy. It excels in handling large datasets with numerous variables and complex interactions between them [23]. CART algorithm is a versatile binary recursive partitioning method for classification and regression tasks [38]. It constructs a tree structure by iteratively dividing data into increasingly homogeneous subsets based on selected features and threshold values; the process continues until stopping conditions, such as reaching maximum depth or a minimum number of samples per leaf node are met [28]. The resulting tree structure can be pruned to prevent overfitting, which improves the generalization capabilities [39]. CART accommodates both categorical and continuous variables, handles multi-class tasks, and provides interpretable decision trees for insights into feature-target connections [40]. GBTA is a powerful machine learning algorithm that enhances prediction accuracy by building an ensemble of decision trees based on the boosting concept [41]. As abovementioned methods, it corrects errors from previous models by iteratively fitting decision trees to residuals and preventing overfitting through regularization [28]. Popular implementations like XGBoost, LightGBM, and GBM, extend GBTA's capabilities, using different optimization techniques [42]. GBTA handles both numerical and categorical variables, demonstrating interpretability and robustness to noisy data; despite its versatility, achieving an optimal performance requires a meticulous hyperparameter tuning [39].

While some studies [30,43,44,45,46] have focused on specific forest attributes or localized areas, a significant gap remains in detailed national-scale mapping of forest attributes. Machine learning (ML) models perform best when training, validation, and testing occur within a well-defined domain. Localized data may introduce uncertainties in species diversity, topography, and forest density, which limit model generalizability. Moreover, ML models face challenges with intra- and inter-class similarity in spectral data, potentially leading to biased predictions. Addressing this requires diverse training datasets encompassing various forest types and topographies. Additionally, hyperparameter tuning may yield good results for a limited domain, but consistency is not guaranteed across broader or changing conditions. Thus, testing ML algorithms at national scales offers insights into their adaptability to varied landscapes [47]. Romania's diverse ecosystems provide an ideal setting for testing the generalizability of ML models [48]. Leveraging this biodiversity can help produce findings that are applicable to other temperate forests worldwide, contributing to improved forest management on both national and global scales [30,49].

The complexity of the model, including the number of parameters and features considered, plays an important role in determining its predictive capabilities [20]. Complex models may be able to capture intricate patterns and relationships within the data but may also be more prone to overfitting, especially when applied to larger spatial extents [50]. On the other hand, simpler models may generalize better but might struggle to capture fine-scale variations in forest attributes [20,43].

Additionally, the choice of spatial resolution affects the performance of machine learning algorithms at larger scales [51]. Coarser spatial resolutions may lead to loss of detail and less variability in the data, potentially affecting the accuracy of predictions [52]; conversely, finer spatial resolutions can provide more detailed information but may also increase computational complexity and resource requirements [49].

Some studies [53,54] advanced our understanding in mapping of forest attributes at a larger scale, such as the national level, but there remains a critical oversight regarding the process of testing on unseen data, which can be provided by independent datasets. This process is essential for assessing the generalizability and robustness of the models [26].

The aim of this study was to model forest characteristics in temperate forests, with a particular focus on Romania, using machine learning algorithms and remotely sensed data. By leveraging the power of machine learning and remote sensing technologies, we aimed to generate comprehensive and detailed maps of forest attributes. Romanian forests were considered representative of temperate forests globally, allowing for the extrapolation of findings to similar forest ecosystems worldwide. The following objectives were setup for this study: i) to compare the performance of three machine learning algorithms—Random Forest (RF), Gradient Boosting Tree Algorithm (GBTA), and Classification and Regression Trees (CART)—in predicting key forest attributes (e.g., standing stock volume, basal area, tree height) based on Sentinel-2 satellite data. This objective supports the selection of the appropriate algorithm by directly comparing different ML approaches, highlighting strengths and weaknesses in their predictions, ii) to evaluate the impact of model complexity on predictive accuracy, elucidating whether increased complexity leads to better performance. This objective aims to determine whether model complexity plays a significant role in improving prediction accuracy, helping us understand when, and if, added complexity is justified, and iii) to assess the effects of spatial resolution on model performance, identifying the optimal resolution for capturing fine-scale forest attributes. This objective provides insights into the types of spatial data required for accurate modelling, guiding decisions regarding data resolution and resource allocation in forest management.

2. Materials and Methods

2.1. Study Area and Field Data

Romanian forests exhibit a high degree of variability, encompassing a wide range of forest types, management practices, and ecological features, which makes them representative for broader European forest conditions. The overall climate of Romania is temperate, (Figure 1), and the mean elevation is 330 m, but it varies widely from 0 to 2544 m a.s.l. [55]. About 18% of the study area holds a rugged terrain (> 15° slope), while the rest is relatively flat (< 5°, 52%), or rolling (5–15°, 30%) [56].

Romanian forests encompass a variety of ecosystems, including mixed broadleaved forests, coniferous stands, and temperate woodlands, each defined by unique structural and compositional traits [57]. The country's topography ranges from lowland plains to mountainous regions, creating diverse microclimates and soil conditions that support species such as beech (Fagus sylvatica), oak (Quercus spp.), and spruce (Picea abies) [58]. Forest management practices range from conservation-focused approaches in protected areas to production forestry, employing sustainable techniques like selective logging and continuous cover forestry to balance biodiversity and ecosystem health [2,59].

These diverse forest conditions make Romania an ideal case study for temperate forest ecosystems, with insights applicable across Europe. At national level, this study relied on ground-based measurements collected from 1,325 circular plots (300–500 m² each) distributed across 10 representative forest areas, sampled from June to September 2022. Data included diameter at breast height, tree height, species, and position, gathered via traditional inventory methods and mobile LiDAR [60]. Metrics like basal area (BA) measured in m²/ha, forest stand stock volume (Vol) measured in m³/ha, average (DBH) measured in cm and dominant tree height (H) measured in m, were calculated using national equations [61] and extrapolated per hectare based on the type of the plot. The national dataset exhibited significant variability, with BA ranging from 0.01 to 95.19 m²/ha (mean: 30.55 m²/ha), DBH averaging 32.19 cm, and Vol at 389.64 m³/ha, reflecting the structural heterogeneity of Romania's forests. For model training at national level, 70% of the data was used, leaving 30% for validation to ensure robust evaluations [62].

An independent testing dataset was collected in Dealul Cetăţii Lempeş area, a temperate forest characterized by diverse habitats and climatic conditions, with a maximum elevation of 704 m. This dataset provided a representative environment for assessing algorithmic predictions. The dataset consists in individual tree measurements to all the trees on an area of 168 hectares (Table D in supplementary material). Forest attribute maps were generated at 10m, 50m, and 100m resolutions by aggregation and extrapolation using QGIS (v3.34.1), enabling detailed analysis of spatial patterns and algorithm performance.

2.2. Remotely Sensed Data and Feature Extraction

The Google Earth Engine (GEE) platform was used to process Sentinel-2 satellite images via JavaScript [63]. This cloud-based solution efficiently handles large datasets, making it ideal for analysing extensive forest areas. The study area included all relevant clusters used for training and the test area, ensuring that models were trained using representative remote sensing data.

Sentinel-2 Level-1C products with a 10-meter spatial resolution were used, offering high-detail images across multiple spectral bands. Specifically, Bands 2 (Blue), 3 (Green), 4 (Red), 8 (NIR), 11, and 12 (SWIR) were deemed suitable for assessing forest attributes like biomass and canopy cover [64].

Essential variables included spectral indices such as NDVI, SAVI, and EVI, as well as slope from SRTM data [65]. These indices were chosen for their effectiveness in quantifying vegetation health and biomass [10,11]. Cloud masking was applied using the Sentinel-2 cloud probability layer, and the imagery was filtered from June to September to minimize seasonal variability. Preprocessing ensured consistency and quality of input data, crucial for improving model performance [66].

All spectral bands required for the indices are merged into a single image, reprojected to the specified coordinate system, and water bodies are masked out to improve accuracy. Next, a feature collection is created, which is essential for both training and validation. In this context, quantitative attributes refer to measurable data extracted from ground-based surveys (e.g., DBH, tree height) and remotely sensed attributes (e.g., NDVI, slope, aspect). These attributes serve as inputs for machine learning, where models learn to extract more abstract features that capture complex patterns not directly observable.

The training-validation dataset combines both ground-based measurements and remote sensing-derived attributes. Machine learning models use this data to learn patterns, validate their predictions, and refine their accuracy. Once validated, these models can reliably estimate forest attributes for new data, supporting robust forest analysis.

2.3. Training and Validation

Once the dataset was prepared with integrated and pre-processed data, we applied machine learning algorithms—CART (Classification and Regression Trees), Random Forest (RF), and Gradient Boosted Trees (GBT)—to model forest attributes. RF and GBT, known for their ability to capture complex relationships through ensemble methods [67,68], were tuned by varying the number of trees to assess the impact of model complexity on prediction accuracy, following hyperparameter optimization strategies recommended in prior studies [20,69]. Specifically, we varied the number of trees from 100 to 2,500 for RF and from 100 to 3,500 for GBT. CART was included for its interpretability despite potential limitations in modelling complex nonlinear relationships. This comparative analysis offers insights into the trade-offs between model complexity, accuracy, and interpretability.

2.4. Performance Assessment

Once the predictive models were developed, we evaluated their performance on the validation dataset using multiple evaluation metrics to ensure a comprehensive assessment. Specifically, we employed the coefficient of determination (R²), root mean squared error (RMSE), relative root mean squared error (rRMSE), and mean absolute error (MAE). Using multiple metrics provides a nuanced understanding of model performance because each captures different aspects of prediction errors. R² measures the proportion of variance in the observed data explained by the model, indicating goodness-of-fit but not reflecting the scale of errors. RMSE calculates the square root of the average squared differences between predicted and observed values, giving higher weight to larger errors. rRMSE normalizes RMSE by the mean of observed values, expressing errors as a percentage, which is useful for comparing models across different scales. MAE computes the average absolute differences between predicted and observed values, providing a measure less sensitive to outliers compared to RMSE. By examining these metrics collectively, we capture both the magnitude and distribution of errors, leading to a more robust evaluation of model accuracy. This approach aligns with recommendations in the literature for comprehensive model assessment.

To statistically assess the influence of different factors on model performance, we conducted an Analysis of Variance (ANOVA). We considered factors such as algorithm type (CART, Random Forest, and Gradient Boosted Trees) and model complexity (number of trees in RF and GBT models, ranging from 100 to 2,500 for RF and 100 to 3,500 for GBT). ANOVA was applied to determine if there were significant differences in performance metrics across different algorithms and complexity levels [70]. Prior to ANOVA, we verified that statistical assumptions were met, and normality of residuals was confirmed using the Shapiro-Wilk test and homogeneity of variances was assessed using Levene's test [71].

Models with the highest R² and lowest RMSE, rRMSE, and MAE were considered to have the best performance. This multi-metric approach ensures that models not only fit the data well but also make precise predictions with minimal errors. The ANOVA results further supported the selection of models with optimal complexity, particularly identifying the number of trees that yielded the most accurate predictions for RF and GBT models.

All analyses were conducted using R version 4.3.2. This evaluation process, combining multiple error metrics and statistical testing, provided a robust framework for selecting the most accurate and reliable models for predicting forest attributes in the study area.

$$\mathrm{R}²=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{{\sum }_{i=1}^n{(y_i-{\overline{y}}_i)}^2}}$$

(1)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{n}}}$$

(2)

$$\mathrm{r}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{RMSE}{\overline{Y}}}\times 100\mathrm{\%}$$

(3)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{n}}{\sum }_{i=0}^n\left|y_i-{\widehat{y}}_i\right|$$

(4)

2.5. Forest Attribute Mapping and Analysis

The trained machine learning models were applied to the satellite imagery of the study area to predict forest attributes at a larger scale. This process involved feeding the satellite imagery into the models and obtaining the predicted values for each pixel containing the selected features. For each attribute, maps were developed at 3 resolutions (10, 50 and 100 m).

To understand how different algorithms predict forest attributes across different resolutions, a detailed analysis was carried out. These maps were compared with the independent dataset that was already prepared for the accuracy assessment (Figure 2). The independent dataset stored as individual tree measurement was assigned geographically on the pixel with the specific resolution of the map (10, 50 or 100m) and extrapolated as stand characteristics at hectare. Then, R², RMSE, MAE and rRMSE were calculated to quantify the agreement between the predicted values and the validation data for each pixel size. This analysis will determine the optimal resolution that enables the capture of fine-scale forest attributes with the highest level of accuracy. And we conducted an ANOVA test to sensitivity of the algorithm’s outputs of the selected models across various spatial resolutions.

3. Results

3.1. Forest Characteristics Modeling Results at National Level

The evaluation of multiple machine learning algorithms for predicting forest stand attributes, summarized in Table 2, highlights significant differences in their performance. Comprehensive wall-to-wall maps were generated for models with the highest prediction accuracy (Figure 3).

The Gradient Boosting Tree Algorithm (GBTA) showed strong predictive capabilities, with R² values increasing from 0.66 (100 trees) to 0.92 (2500 trees) for stand average volume (Vol) estimation, indicating improved performance with increased model complexity. Similar trends were observed for other attributes (Figure 5). GBTA predicted higher volumes in regions similar to Random Forest (RF) but with notable differences, particularly in the central and northeastern areas (Figure 3).

RF, on the other hand, performed well, explaining over 80% of the variance in the target variables. This regression line performed well despite being slightly less steep than GBTA with 2500 trees as shown in Figure 4. Consequently, the R2 value was 94.09% and the RMSE was approximately 3.56 m2/ha. In Figure 3, The model predicts higher forest Vol, BA, DBH, and H predominantly in the central and northwestern regions of Romania, with lower values generally toward the borders.

In BA estimations, CART outperformed all other predictors with its high R2 value of 94.09% and low RMSE of 3.56. Nevertheless, its accuracy did not extend uniformly across all forest attributes, as it failed to consistently outperform RF and GBTA. Thus, while it excelled in explaining the variance in BA predictions, its performance exhibited variability across different scenarios. Visually, CART presents a more homogeneous distribution with fewer areas of high BA concentration compared to other models.

Table 1. Performance of machine learning algorithms in the validation datasets attributes and scenarios.

Attribute	Model	N0 of trees	RMSE	MAE	rRMSE(%)	R2
Vol (m3/ha)	RF	100	107.461	-2.420	0.284	0.810
		1000	107.727	-1.217	0.285	0.810
		3500	107.643	-1.013	0.285	0.810
	GBTA	100	178.085	17.077	0.471	0.660
		1000	89.769	2.105	0.237	0.860
		2500	64.507	1.499	0.171	0.930
	CART	1000	132.663	0.000	0.351	0.630
BA (m2/ha)	RF	100	6.617	-0.020	0.223	0.830
		1000	6.594	-0.047	0.222	0.830
		3500	6.585	-0.046	0.222	0.830
	GBTA	100	11.520	1.039	0.388	0.710
		1000	5.286	-0.177	0.178	0.889
		2500	3.809	-0.113	0.128	0.940
	CART	1000	3.564	0.000	0.120	0.941
DBH (cm)	RF	100	6.402	-0.005	0.205	0.816
		1000	6.311	-0.029	0.202	0.826
		3500	6.294	-0.034	0.201	0.827
	GBTA	100	10.768	0.923	0.345	0.708
		1000	5.154	-0.061	0.165	0.874
		2500	3.616	3.616	-0.044	0.936
	CART	1000	7.904	0.000	0.253	0.653
H (m)	RF	100	3.207	-0.022	0.135	0.839
		1000	3.198	-0.009	0.134	0.845
		3500	3.192	-0.014	0.134	0.845
	GBTA	100	2.589	-0.290	0.109	0.750
		1000	2.589	-0.290	0.109	0.891
		2500	1.868	-0.199	0.078	0.941
	CART	1000	4.234	0.000	0.178	0.665

Table 1. Performance of machine learning algorithms in the validation datasets attributes and scenarios.

Attribute	Model	N0 of trees	RMSE	MAE	rRMSE(%)	R2
Vol (m3/ha)	RF	100	107.461	-2.420	0.284	0.810
		1000	107.727	-1.217	0.285	0.810
		3500	107.643	-1.013	0.285	0.810
	GBTA	100	178.085	17.077	0.471	0.660
		1000	89.769	2.105	0.237	0.860
		2500	64.507	1.499	0.171	0.930
	CART	1000	132.663	0.000	0.351	0.630
BA (m2/ha)	RF	100	6.617	-0.020	0.223	0.830
		1000	6.594	-0.047	0.222	0.830
		3500	6.585	-0.046	0.222	0.830
	GBTA	100	11.520	1.039	0.388	0.710
		1000	5.286	-0.177	0.178	0.889
		2500	3.809	-0.113	0.128	0.940
	CART	1000	3.564	0.000	0.120	0.941
DBH (cm)	RF	100	6.402	-0.005	0.205	0.816
		1000	6.311	-0.029	0.202	0.826
		3500	6.294	-0.034	0.201	0.827
	GBTA	100	10.768	0.923	0.345	0.708
		1000	5.154	-0.061	0.165	0.874
		2500	3.616	3.616	-0.044	0.936
	CART	1000	7.904	0.000	0.253	0.653
H (m)	RF	100	3.207	-0.022	0.135	0.839
		1000	3.198	-0.009	0.134	0.845
		3500	3.192	-0.014	0.134	0.845
	GBTA	100	2.589	-0.290	0.109	0.750
		1000	2.589	-0.290	0.109	0.891
		2500	1.868	-0.199	0.078	0.941
	CART	1000	4.234	0.000	0.178	0.665

Figure 3. Forest characteristics modelling results at national level.

Figure 3. Forest characteristics modelling results at national level.

Figure 4. Scatterplots for the RF regression model of forest stand attributes: (a) basal area (b) tree volume (Vol) (c) diameter at breast height (DBH) (d) tree height (H).

Figure 4. Scatterplots for the RF regression model of forest stand attributes: (a) basal area (b) tree volume (Vol) (c) diameter at breast height (DBH) (d) tree height (H).

Figure 5. Scatterplots for the GBTA regression model of forest stand attributes: (a) basal area (b) tree volume (Vol) (c) diameter at breast height (DBH) (d) tree height (H).

Figure 5. Scatterplots for the GBTA regression model of forest stand attributes: (a) basal area (b) tree volume (Vol) (c) diameter at breast height (DBH) (d) tree height (H).

Figure 6. Scatterplots for the CART regression model of forest attributes: (a) basal area (b) tree volume (Vol) (c) diameter at breast height (DBH) (d) tree height (H).

Figure 6. Scatterplots for the CART regression model of forest attributes: (a) basal area (b) tree volume (Vol) (c) diameter at breast height (DBH) (d) tree height (H).

3.2. Model Performance Assessment with Independent Validation Data in the Test Area

The assessment of prediction accuracy and reliability through comparison with independent ground truth data in Lempes test area indicates that, RF and GBTA algorithms with a maximum number of suggested trees demonstrated high performance in predicting various forest attributes during the evaluation with initial validation data. The CART algorithm exhibited reliable outcomes specifically in predicting BA. Based on these findings, these three models were selected for further analysis. This selection process was crucial to ensure the robustness and generalizability of the chosen models.

3.2.1. Visual Assessment of the Models in Mapping Forest Characteristics

The detailed maps presented in Figure 7, Figure 8, Figure 9 and Figure 10 illustrate the predictive modelling of forest attributes across the entire test area. These intricate cartographic representations provide a granular perspective on model predictions at a spatial resolution of 10 meters. The three models show varying degrees of accuracy in their estimations compared to observed data.

GBTA consistently provides predictions that closely align with the actual data, particularly excelling in capturing the upper range of values such as taller tree heights, larger DBH, higher volumes, and BA. RF generally achieves accuracy comparable to GBTA but tends to produce more moderate predictions. It often smooths out extremes and may slightly underpredict in areas with exceptionally large trees. CART, on the other hand, tends to underpredict, especially in regions with higher observed values, showing a more conservative approach, particularly in predicting larger trees in terms of H, DBH, and Vol. Overall, GBTA is the most accurate visually for environments with high variability, RF offers reliable but more moderate predictions, and CART is best suited for scenarios where lower, conservative estimates are preferable.

3.2.2. Evaluation of Predictive Accuracy in Forest Attribute Estimation Across Different Resolutions

The accuracy of these algorithms in predicting the independent dataset, when varying the pixel size from low to high are presented in Table 3. This comprehensive assessment allowed an optimal determination of the balance between detail and precision for optimal prediction.

At the 10-meter resolution, the accuracy of the models is generally low. This finer resolution tends to introduce more variability and noise into the data, which the models struggle to handle effectively. Consequently, the coefficient of determination values are lower, suggesting that the models account for less variance in the data. Furthermore, the error metrics are higher, indicating reduced predictive performance.

When examining DBH and H, both RF and GBTA algorithms demonstrate significant improvement in R2 at 100 meters (about 55% the dataset was accurately predicted) with the lowest RMSE and rRMSE values, suggesting that higher pixel sizes lead to better predictive accuracy for these attributes. The CART algorithm follows a similar trend, with the best performance observed at 100 meters for both DBH and Height with 41.7%. This proves that for predicting DBH and Height, higher resolutions (100 meters) generally provide more accurate and precise results across all three algorithms, particularly for RF and GBTA.

For Vol, the RF and CART algorithm shows slight improvement in R2 at 100 meters, but this comes with significantly higher RMSE and rRMSE. In contrast, the GBTA algorithm demonstrates superior performance, with notable improvements in accuracy and decreasing errors at both 50-meter and 100-meter resolutions. For BA, the optimal accuracy and precision was at 50 m resolution for RF and GBTA algorithm, with a significant improvement in error metrics. The CART algorithm performs best at 10 meters but shows a decline in accuracy at higher resolutions. This indicates that the models struggle particularly with predicting Vol and BA effectively even when considering larger resolution.

Figure 11. Scatterplots for the models’ performance with the test area for BA prediction under different resolution.

Figure 11. Scatterplots for the models’ performance with the test area for BA prediction under different resolution.

Figure 12. Scatterplots for the models’ performance with the test area for H prediction under different resolution.

Figure 12. Scatterplots for the models’ performance with the test area for H prediction under different resolution.

Figure 13. Scatterplots for the models’ performance with the test area for DBH prediction under different resolution.

Figure 13. Scatterplots for the models’ performance with the test area for DBH prediction under different resolution.

Figure 14. Scatterplots for the models’ performance with the test area for Vol prediction under different resolution.

Figure 14. Scatterplots for the models’ performance with the test area for Vol prediction under different resolution.

4. Discussion

Forests are essential ecosystems that sustain biodiversity, regulate climate, and provide critical resources such as timber and carbon sequestration capabilities. The increasing demand for sustainable forest management requires efficient and scalable tools for monitoring key forest attributes, including basal area (BA), standing stock volume (Vol), diameter at breast height (DBH), and tree height (H). Traditional ground-based measurements, while accurate, are resource-intensive and often impractical for large-scale assessments. This study leverages remote sensing technologies and machine learning (ML) algorithms—Random Forest (RF), Gradient Boosting Tree Algorithm (GBTA), and Classification and Regression Trees (CART)—to address these limitations, offering a modern, scalable solution for predicting forest attributes in Romania's diverse temperate forest ecosystems.

The main objectives were to evaluate the performance of these ML algorithms in predicting key forest attributes, assess the impact of model complexity on predictive accuracy, and determine the optimal spatial resolution for modeling fine-scale forest characteristics. These objectives align with the overarching goal of enhancing forest management practices globally by providing detailed and accurate forest attribute maps.

4.1. Comparative Algorithm Performance

The study revealed notable differences in the performance of RF, GBTA, and CART across various forest attributes. GBTA emerged as the most robust algorithm, particularly for complex and variable attributes such as Vol and BA. At its optimal complexity (2,500 trees), GBTA achieved an R² of 0.92 for standing volume, outperforming other algorithms in terms of both accuracy and error metrics. This aligns with findings from similar studies where GBTA was favored for its ability to model intricate relationships in forest ecosystems [43,72,73].

RF demonstrated strong performance, explaining over 80% of the variance for most attributes, making it a reliable choice for large-scale applications. However, its tendency to underpredict in areas with extreme values was observed, as reported in prior studies [34,68]. The algorithm's moderate computational requirements and robustness to overfitting underscore its practicality for forest management applications.

CART, known for its interpretability, performed best in BA prediction, achieving the lowest RMSE values. Nevertheless, its conservative predictions for attributes such as DBH and H highlight its limitations in handling complex, nonlinear relationships. This corroborates findings from studies suggesting CART’s suitability for scenarios requiring simple and transparent decision-making frameworks [38,39].

4.2. Influence of Model Complexity and Impact of Spatial Resolution

A key objective of this study was to assess whether increased model complexity enhances predictive accuracy. Results confirmed that while higher complexity generally improved performance, diminishing returns were observed beyond a certain threshold. For instance, GBTA's accuracy plateaued as tree numbers exceeded 2,500, emphasizing the importance of balanced hyperparameter optimization. These findings align with previous research, which highlights the trade-offs between model complexity and overfitting risks [25].

RF displayed resilience to overfitting, achieving consistent accuracy across different complexity levels. This characteristic makes it particularly valuable for large-scale forest assessments, where overfitting poses a significant challenge [35,36]. In contrast, CART exhibited greater sensitivity to model complexity, reinforcing its role as a simpler yet less flexible algorithm for specific applications.

The impact of spatial resolution on model performance was a pivotal focus of this study. Fine resolutions (10 m) captured detailed variability but introduced noise, reducing predictive accuracy. Intermediate resolutions (50 m) offered a balance between spatial detail and model performance, proving optimal for attributes like BA and H. Coarser resolutions (100 m) improved the accuracy of DBH predictions but sacrificed detail, particularly in heterogeneous landscapes.

These findings are consistent with studies that emphasize the trade-offs between resolution and accuracy in remote sensing applications [49,51]. The variability in optimal resolution across attributes underscores the need for tailored approaches based on specific management objectives. For instance, finer resolutions may be preferable for biodiversity assessments, while coarser resolutions may suffice for large-scale carbon stock estimation.

4.3. Broader Applicability of Results and Contributions to Forest Management

Romanian temperate forests, characterized by diverse species, management practices, and topographic conditions, serve as a representative case study for broader European and global contexts. The applicability of the results extends to other temperate forest ecosystems, where similar challenges in monitoring and managing forest attributes persist. The integration of Sentinel-2 data with ML algorithms offers a replicable framework for sustainable forest management across regions with varying ecological and management conditions.

The robust performance of GBTA and RF in this study aligns with global efforts to leverage machine learning for forest monitoring, highlighting their potential for scalable and accurate applications in diverse forest types [23,74]. Additionally, the adaptability of these algorithms to varying resolutions and attribute complexities enhances their utility for addressing regional and global forest management challenges.

By meeting the study's objectives, this research contributes significantly to the field of forest management. The ability to generate accurate, large-scale maps of forest attributes supports decision-making processes for timber harvesting, carbon accounting, and biodiversity conservation. The integration of advanced ML techniques with satellite data reduces the reliance on resource-intensive ground surveys, enabling cost-effective and scalable monitoring solutions.

The findings also emphasize the importance of selecting appropriate algorithms and resolutions based on management priorities. For example, GBTA is suitable for detailed assessments requiring high accuracy, while RF offers a practical alternative for broader applications. This flexibility ensures that forest managers can adapt these tools to specific needs, promoting efficient and sustainable resource management.

5. Conclusions

This study demonstrates the transformative potential of machine learning and remote sensing in forest management. By addressing the limitations of traditional methods, it provides a scalable, accurate, and efficient framework for monitoring forest attributes. The use of Romania’s forests as a representative case study highlights the global relevance of these findings, offering insights applicable to temperate forest ecosystems worldwide. Ultimately, the integration of advanced technologies into forest management practices represents a significant step toward achieving sustainability goals, balancing conservation and utilization to meet the needs of present and future generations.