Mapping of Forest Species Using Sentinel-2A Images in the Alentejo and Algarve Regions, Portugal

Abstract

Land use and land cover (LULC) studies, particularly those focused on mapping forest species using Sentinel-2 (S2A) data, face challenges in delineating and identifying areas of heterogeneous forest components with spectral similarity at the canopy level. In this context, the main objective of this study was to compare and analyze the feasibility of two classification algorithms, K-Nearest Neighbor (KNN) and Random Forest (RF), with S2A data for mapping forest cover in the southern regions of Portugal, using tools with a free, open-source, accessible, and easy-to-use interface. Sentinel-2A data from summer 2019 provided 26 independent variables at 10 m spatial resolution for the analysis. Nine object-based LULC categories were distinguished, including five forest species (Quercus suber, Quercus rotundifolia, Eucalyptus spp., Pinus pinaster, and Pinus pinea), and four non-forest classes. Orfeo ToolBox (OTB) proved to be a reliable and powerful tool for the classification process. The best results were achieved using the RF algorithm in all regions, where it reached the highest accuracy values in Alentejo Central region (OA = 92.16% and K = 0.91). The use of open-source tools has enabled high-resolution mapping of forest species in the Mediterranean, democratizing access to research and monitoring.

1. Introduction

Mapping changes in land use and land cover (LULC) is essential for understanding ecosystem dynamics and ensuring sustainable natural resource management. Several methods and techniques have been used, among which are those linked to remote sensing, especially with Sentinel-2 data. Mapping LULC is a challenge when landscape has large variability and fragmentation, such as in the Mediterranean region [1]. The main land use classes in continental Portugal are forest (36%), woods and pastures (31%), agricultural areas (24%), urban use (5%), unproductive (2%), and inland waters (2%) [2]. Considering the aforementioned, it is pivotal to study the potentiality of remote sensing data to map Mediterranean ecosystem. This will enable us to map forest species distribution. For these reasons, maps serve as an intermediate step in more complex analyses such as landscape modeling, estimating forest biomass, and environmental monitoring [3,4]. In addition to being a region prone to forest degradation and wildfire, as is the case for many Mediterranean forests [5,6], the forest species’ identification and spatial distribution are essential to monitoring biodiversity, climate change, and forest management.

Remote sensing has undergone significant changes in recent years, resulting in a substantial increase in both the temporal frequency and spatial resolution of satellite imagery, as well as improved data accessibility. Several satellite types are available on open-access platforms, including high-resolution satellite imaging systems (Friedl et al. [7]), and several non-commercial software packages have image analysis tools [8,9,10].

Sentinel-2A (S2A) is a multispectral sensor (MSI) which has a five-day revisit time, high spatial resolution and its data are open sourced [11]. It is composed of 13 spectral bands, with a wavelength between 443 and 2190 nm, ranging from the visible spectrum, vegetation red-edge bands (Red-edge), near-infrared (NIR), narrow infrared (Narrow NIR), and shortwave infrared (SWIR) [11]. Depending on the band, three spatial resolutions are available: 10 m (B2, B3, B4 (V), and B8 (NIR)), 20 m (B5, B6, B7, (Red-edge), B8a (Narrow NIR), B11 and B12 (SWIR)), and 60 m (B1, B9, and B10). The S2A product can be acquired with corrected reflectance images at the base of the atmosphere and the closest ground measurements (Bottom Of Atmosphere—BOA) [12]. S2A when compared to Landsat mission and MODIS data have the advantage of having more spectral bands and higher spatial resolution, respectively [13].

In addition to spectral information from Sentinel-2 bands, vegetation and texture indices are widely used in remote sensing to improve LULC classification and forest mapping [14,15]. These indices provide complementary information to the spectral band wavelengths, enabling more accurate identification and mapping of Mediterranean vegetation types and different types of land cover. Vegetation indices are a type of spectral indices that specifically target the reflectance properties of vegetation. These indices are calculated using the red and near-infrared bands of a satellite image and can be used to estimate vegetation biomass, health, and productivity. In particular, vegetation indices, such as Green Normalized Difference Vegetation Index (GNDVI), Soil-Adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI), Normalized Difference Infrared Index (NDII), Normalized Difference Vegetation Index red-edge 1 narrow (NDRE 1), Normalized Difference Vegetation Index red-edge 2 narrow (NDRE 2), and Red-edge Chlorophyll Index (CI), can be used to discriminate different vegetation types [16,17]. Their advantage is related to being sensitive to patterns, especially in semi-arid environments [18].

Texture indices refer to the spatial arrangement of different land cover types within an image. They are calculated by analyzing the variation in pixel values within a local neighborhood of the image. These indices can be used to identify patterns in the landscape that are not easily discernible from color imagery alone [19]. For LULC classification purposes, the texture measurements of a Gray Level Co-occurrence Matrix (GLCM), proposed by Haralick et al. [20], provide further information, which allows us to classify heterogeneous land uses that spectral information is not able to differentiate [19], which is why it is a widely used method. The GLCM method analyzes the existing co-occurrences between each pair of pixels in an image; that is, it analyzes the spatial relationship between a set of pixels (given a distance and different angles) [20,21,22]. GLCM, being not related to spectral data, frequently attains higher image classification accuracy [1,23,24]. The GLCM texture index can be used to identify areas with high or low levels of spatial heterogeneity, which can be useful for distinguishing between different forest types [25,26]. Principal component analysis (PCA) is one of the commonly used methods to calculate GLCM textures from spectral bands or vegetation indices [18,22], with the goal of simplifying and reducing the number of explanatory variables and the correlation between them [27,28].

In the supervised classification method, training selection and classification methods are also crucial steps in accurately mapping forest classes using satellite data. The quality of the training samples and the appropriateness of the classification method used determine the accuracy of the final forest classification map [29,30]. In recent decades, methods based on machine learning, non-parametric methods, have been frequently applied in remote sensing [31,32], such as K-Nearest Neighbor (kNN) [33] and Random Forest (RF) [34]. kNN is a non-parametric learning method proposed by Wettschereck et al. [35]. It does not make assumptions about the training dataset, since it is independent of the data spectral distribution [36]. In this way, kNN is based on a calibration dataset, where it finds a set of k training samples closest to the unknown samples, frequently based on the Euclidean distance function [36,37,38,39]. Random Forest (RF), proposed by [34], divides datasets into two subsets (i.e., training and validation). The method is based on the definition of decision trees with random bootstrapping of training data, which, inversely to the decision tree (DT) method, does not have overfitting problems [40,41]. A random subset of predictors is used at each node to divide the tree into decision trees. The output is the mean of all decision trees, based on the majority of “votes” per class (unweighted) [42,43].

These methods perform well in data categorization where input complexity is high, and they are widely accepted in LULC mapping [4,43,44,45,46]. Several studies compared different classification algorithms to classify LULC with Sentinel-2 data. The RF algorithm was compared with other commonly used classification methods, including Maximum Likelihood (ML), support vector machine (SVM), and KNN [30,47,48]. For all studies, the RF algorithm presents better results in terms of overall accuracy, kappa coefficient, and user accuracy. The authors discussed the importance of selecting appropriate training samples and choosing the appropriate classification method for a particular forest type to ensure accurate classification. More information on supervised classification methods can be found in [49].

In particular, the Mediterranean ecosystem presents a unique set of challenges due to its complex vegetation patterns, topography, and land use changes. In these ecosystems, where there is great structural and botanical diversity with high spectral variability [50], supervised classification methods based on machine learning algorithms have attracted attention for being able to deal with large volumes of complex and non-parametric data [43,51]. For Mediterranean systems, some studies have been developed to obtain LULC maps with Sentinel-2 data, in Mediterranean ecosystems in Greece [52], Spain [53,54], Tunisia [55], and Italy [46]. These studies compared RF with other classification algorithms such as Convolutional Neural Networks (CNNs), support vector machines (SVMs), Multilayer Perceptron, Maximum Likelihood, K-Dimensional Trees K-Nearest Neighbors (KDTree-KNNs), Minimum Distance Classification (MDC), and classification and regression tree (CART). The results for these studies showed that RF achieved the highest classification accuracy, with overall accuracy higher than 88%. The RF algorithm is a popular choice for remote sensing applications and is the main objective due to its ability to handle large datasets, reduce overfitting, and provide accurate and interpretable results.

This study presents a LULC mapping approach entirely developed in an open-source environment, aiming to explore its potential for collecting sampling data and classifying high-resolution images from Sentinel-2. In the supervised classification, the Orfeo ToolBox (OTB) version 6.6.1 is an application available in the open-source since software QGIS version 3.8 through a Python plugin [56]. The OTB has specific characteristics suitable for spatial data analysis and high-resolution image processing [57]. This open-source project receives regular updates from the open-source geospatial community, ensuring its relevance in studies such as vegetation monitoring [58]. Its compatibility with various remote sensing data formats, including Sentinel-2 images, increases its versatility and applicability in supervised classification [58]. A significant advantage of OTB is its accessibility through the software QGIS plugin, allowing researchers and professionals with limited programming knowledge to use advanced supervised classification tools. OTB is composed of several classification algorithms, such as Random Forests and K-Nearest Neighbors (KNNs), and processes large data at scale [56]. Furthermore, this approach enabled the exploitation of the strengths of supervised classification without the need for advanced programming knowledge, which optimized the acquisition of a high-precision forest cover map.

The significant contribution of this research lies in the proposed workflow for analyzing different image datasets from Sentinel-2 satellite and variables with vegetation and texture indices, aiming to enhance the classification accuracy of forest species in southern Portugal. This study addresses the application of machine learning to detect the main challenges in generating high-precision, large-scale classified LULC maps in highly complex forested landscapes. These landscapes are characterized by mixed forest compositions and a heterogeneous, fragmented, and dispersed spatial pattern. The methodological approach focuses on demonstrating the possibility of creating precise maps using open-source tools and data, making the methodology democratically accessible. The proposal aims to achieve scientific reproducibility in LULC mapping studies, enabling access for university students, professionals, and GIS enthusiasts with basic knowledge, thus promoting the dissemination of geospatial knowledge. To enhance the accessibility of analysis, classification, and mapping processes for different users with diverse backgrounds and varying software and hardware availability, the main objectives of this study were as follows: (i) to produce a mapping of forest species cover in the Alentejo and Algarve regions, southern Portugal, using remote sensing data derived from the Sentinel-2 (S2A) satellite; (ii) to develop a methodology and evaluate and compare the feasibility of KNN and RF classification algorithms in S2A; (iii) to use free and open-source data and image analysis tools, focusing on accessibility for users without advanced programming knowledge, offering a user-friendly interface and the ability to process large datasets; (iv) to use the open-source QGIS and SNAP software and the OTB plugin, an image processing library, to classify and analyze Sentinel-2 (S2A) data.

2. Materials and Methods

2.1. Study Area

The study area covers NUT II (NUT—Nomenclature of Territorial Units for Statistical Purposes), Alentejo (AL) and Algarve (AG) [59]. The Alentejo is divided into four NUT III, namely Alto Alentejo (AA), Alentejo Central (AC), Alentejo Litoral (AL), and Baixo Alentejo (BA) (Figure 1) [59]. These regions are located in the southern mainland of Portugal, with a total territorial extension of 36 602 km², with a Mediterranean climate, with hot and dry summers (June to September) and cold and rainy winters (October to January). Total forest area is about 1,270,000 ha, and the most representative species in terms of area are Quercus suber, Quercus rotundifolia, Eucalyptus spp., Pinus pinea, and Pinus pinaster [2].

Quercus suber, Quercus rotundifolia, and Pinus pinea occupy the largest share of the forest area and are managed as agroforestry systems [60], which have low density, heterogeneous spatial distribution, in pure stands or mixed with other species, and usually shrub and herbaceous species in the understory [61,62,63,64]. Eucalyptus spp. and Pinus pinaster are managed as forest systems for timber [65,66] and have higher density and more homogenous spatial distribution than the former [64].

2.2. Data Processing and Dataset

The methodology for the satellite image processing is outlined in Figure 2, which includes the main following steps: (1) selection of Sentinel-2 image processing, calculation of vegetation indices, and calculation of the principal component analysis (PCA) and the GLCM textures (cf. Section 2.3); (2) definition of training and test samples; (3) application of the supervised classification algorithms (cf., Section 2.4.1); and (4) accuracy assessment of different classifiers on a test set using the confusion matrix (cf., Section 2.4.2).

2.3. Spectral Data

Copernicus Open Access Hub platform [11] was used to acquire S2A images. The image product used was the atmospherically corrected surface reflectance, Level 2A, derived from the associated Level 1C product, radiometrically corrected imagery. To cover the entire study area (Figure 1 right), six S2A images were used (approximately 1 GB each file) with the same date acquisition (2 September 2019) and cloud filtering of less than 10%. In this study, 10 S2A Multispectral (MSI) reflectance bands were used (B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12); the mosaics per reflectance band were set in European Terrestrial Reference System 1989 (ETRS 1989). The summer season was selected because, in the Mediterranean environment, the probability to find images with a lower percentage of cloud cover and contrast between the spectral signature of the different forest species and the understory vegetation is larger [18,67]. To obtain all bands with the same spatial resolution, spatial upscaling using the nearest neighbor method was applied to convert the 20m bands (5, 6, 7, 8A, 11, and 12) into 10m spatial resolution [68].

Seven vegetation indices (VIs) and nine texture indices (Gray Level Co-occurrence Matrix—GLCM) were implemented to promote spectral discrimination of forest classes (Table 1). This approach became necessary due to the intrinsic limitations of broad spectral bands, such as their susceptibility to atmospheric interference and the resulting similar spectral reflectance patterns in different vegetation types [50,69]. Traditional VIs, as well as some more recent VIs utilizing red-edge spectral bands available in Sentinel-2 data, were employed to contribute to better separability of land use and land cover (LULC) classes, with an emphasis on vegetation classes. The processing of Sentinel imagery and the calculation of vegetation indices (VIs) were performed using QGIS 3.16.0-Hannover software.

Principal component analysis was applied to the set of 10 Sentinel-2A spectral bands, following recommendations from various studies [27,70,71]. In this study, the first three principal components were selected for subsequent analysis, as they explained more than 97.10% of the spectral variation of the dataset for all regions (Table 2). This approach is based on the premise, corroborated by previous research [1,22,72], that the first three principal components retain most of the relevant information for discriminating land use and land cover classes.

From these three principal components, each GLCM texture measure (Average—MEA, Variance—VAR, and Correlation—COR) was calculated using a 9 × 9 pixel moving window in the 0°, 45°, 90°, and 135° directions of 32 gray levels, with the variogram method [73]. The size of the moving window depends on the land use characteristics and the diversity of the bands [50,74]. Szantoi et al.’s [75] window size was adapted due to its ability to evaluate heterogeneous regions, regardless of the size of the training features. GLCM textures were generated using SNAP (Sentinel Application Platform) software version 7 [75]. In this study, the selection of vegetation and texture indices was based on their demonstrated performance in previous research [76,77,78]. In total, the input dataset for this study comprised 26 independent variables.

Table 1. Description and formula of the indices.

Types	Descriptions		Equations
Vegetation Indices * (VIs)	GNDVI [79]	Green Normalized Difference Vegetation Index	${\displaystyle \frac{B8-B3}{B8+B3}}$
	SAVI [80]	Soil-Adjusted Vegetation Index	$1.5\times {\displaystyle \frac{B8-B4}{8\times \left(B8+B4+0.5\right)}}$
	NDII [81]	Normalized Difference Infrared Index	${\displaystyle \frac{B8-B11}{B8+B11}}$
	EVI [82]	Enhanced Vegetation Index	$2.5\times \left({\displaystyle \frac{B8-B4}{\left(B8-6\times B4-\left(B4-7.5\times B2\right)+1\right)}}\right)$
	NDRE 1 [83]	Normalized Difference Red Edge Index 1	${\displaystyle \frac{B8A-B5}{B8A+B5}}$
	NDRE 2 [83]	Normalized Difference Red Edge Index 2	${\displaystyle \frac{B8A-B7}{B8A+B7}}$
	CI [84]	Red-edge Chlorophyll Index	${\displaystyle \frac{B6}{B5}}$
Textures **	Gray-level co-occurrence matrix (GLCM) [20]	MEA	${\sum }_{i,j=0}^{N-1}i P_{i,j}$
		VAR	${\sum }_{i,j=0}^{N-1}i{P}_{i,j} {(i-\mu )}^2$
		COR	${\displaystyle \frac{{\sum }_{i,j=0}^{N-1}i{P}_{i,j}-{\mu }_x {\mu }_y}{{\sigma }_x {\sigma }_y}}$

* where the central wavelength (µm): B2 (490), B3 (560), B4 (665), B5 (705), B6 (740), B7 (783), B8 (842), B8A (865), B11 (1610), and B12 (2100). ** where P(_i,j) is a normalized gray-tone spatial dependence matrix such that SUM(i,j = 0, N − 1) (P(_i,j)) = 1; i and j represent the rows and columns, respectively, for the measures of Mean, Variance and Correlation; μ is the mean, for the Variance textural measure; and N is the number of distinct gray levels in the quantized image; μ_x. μ_y, σ_x, and σ_y are the means and standard deviations of p_x (sum of each row in co-occurrence matrix) and p_y (sum of each column in the co-occurrence matrix), respectively.

Table 1. Description and formula of the indices.

Types	Descriptions		Equations
Vegetation Indices * (VIs)	GNDVI [79]	Green Normalized Difference Vegetation Index	${\displaystyle \frac{B8-B3}{B8+B3}}$
	SAVI [80]	Soil-Adjusted Vegetation Index	$1.5\times {\displaystyle \frac{B8-B4}{8\times \left(B8+B4+0.5\right)}}$
	NDII [81]	Normalized Difference Infrared Index	${\displaystyle \frac{B8-B11}{B8+B11}}$
	EVI [82]	Enhanced Vegetation Index	$2.5\times \left({\displaystyle \frac{B8-B4}{\left(B8-6\times B4-\left(B4-7.5\times B2\right)+1\right)}}\right)$
	NDRE 1 [83]	Normalized Difference Red Edge Index 1	${\displaystyle \frac{B8A-B5}{B8A+B5}}$
	NDRE 2 [83]	Normalized Difference Red Edge Index 2	${\displaystyle \frac{B8A-B7}{B8A+B7}}$
	CI [84]	Red-edge Chlorophyll Index	${\displaystyle \frac{B6}{B5}}$
Textures **	Gray-level co-occurrence matrix (GLCM) [20]	MEA	${\sum }_{i,j=0}^{N-1}i P_{i,j}$
		VAR	${\sum }_{i,j=0}^{N-1}i{P}_{i,j} {(i-\mu )}^2$
		COR	${\displaystyle \frac{{\sum }_{i,j=0}^{N-1}i{P}_{i,j}-{\mu }_x {\mu }_y}{{\sigma }_x {\sigma }_y}}$

* where the central wavelength (µm): B2 (490), B3 (560), B4 (665), B5 (705), B6 (740), B7 (783), B8 (842), B8A (865), B11 (1610), and B12 (2100). ** where P(_i,j) is a normalized gray-tone spatial dependence matrix such that SUM(i,j = 0, N − 1) (P(_i,j)) = 1; i and j represent the rows and columns, respectively, for the measures of Mean, Variance and Correlation; μ is the mean, for the Variance textural measure; and N is the number of distinct gray levels in the quantized image; μ_x. μ_y, σ_x, and σ_y are the means and standard deviations of p_x (sum of each row in co-occurrence matrix) and p_y (sum of each column in the co-occurrence matrix), respectively.

Table 2. Cumulative values of principal component (PC) eigenvalues.

NUTIII Regions	AA	AC	AL	BA	AG
PC1	77.69%	96.87%	75.64%	81.60%	95.77%
PC1 + PC2	93.16%	98.81%	92.99%	93.15%	98.24%
PC1 + PC2 + PC3	97.18%	99.78%	97.42%	97.10%	99.45%

where AA—Alto Alentejo, AC—Alentejo Central, AL—Alentejo Litoral, and BA—Baixo Alentejo.

Table 2. Cumulative values of principal component (PC) eigenvalues.

NUTIII Regions	AA	AC	AL	BA	AG
PC1	77.69%	96.87%	75.64%	81.60%	95.77%
PC1 + PC2	93.16%	98.81%	92.99%	93.15%	98.24%
PC1 + PC2 + PC3	97.18%	99.78%	97.42%	97.10%	99.45%

where AA—Alto Alentejo, AC—Alentejo Central, AL—Alentejo Litoral, and BA—Baixo Alentejo.

2.4. Supervised Classification

The LULC classes, in the study area, defined by polygons (Regions of Interest—ROIs features) have a uniform distribution [85]. The knowledge of the study area is fundamental to have representative training and validation data samples, for the process of image classification [9]. Thus, for the delimitation of the ROI features, per land use class, we utilized reference data, such as high-resolution satellite images from Google Satellite [86] and the Land Use and Occupation Map (COS) of 2018 from Portugal [87]. When marking the ROI samples, the MetaSearch plugin integrated into QGIS simplifies searching in metadata catalogs compatible with the CSW (Catalog Service for the Web) standard from the Open Geospatial Consortium (OGC) [86]. With a user-friendly interface, the MetaSearch plugin enables access to various services, such as Google Earth Images, facilitating the discovery, navigation, and consultation of geographic metadata [86].

In this study, we defined nine LULC classes, five forest and four non-forest. The five forest classes correspond to five forest species, namely Eucalyptus spp., Pinus pinea, Pinus pinaster, Quercus rotundifolia, and Quercus suber. The four non-forest classes are as follows: water surfaces, bare soil and artificial surfaces, agricultural surface, and shrub surface. We chose to include the bare soil class and urban areas in a single class (bare soil and artificial surfaces). In this study, bare soil and urban areas, as well as agricultural surface classes, were consolidated into single classes, as the primary focus of the research was the classification of major forest covers. During the summer period, the consolidation of agricultural species into a single class was carried out because they exhibit phenological cycles that differ from forest species, resulting in areas of high temporal variability when grouped, which contrasts with the relative stability of forested areas [88]. However, the spectral resolution constraints of Sentinel-2 imagery can present challenges in consistently distinguishing between different crop types [88]. Furthermore, both exposed surfaces bare soil, and urban areas tend to exhibit elevated reflectance values due to low soil moisture content [89] and the presence of artificial materials such as concrete, asphalt, and roofs [90]. This spectral similarity presents challenges in discriminating between these classes, thus supporting their aggregation into a unified category. In this sense, these aggregations reduced the computational complexity of the classification process, allowing for a more efficient and robust analysis, particularly for large-scale studies. In this study, supervised classification was applied using the OTB plugin. The independent stratified random sampling method was used; 70% of the ROIs were used for training the classifiers, while 30% were used for validation and accuracy assessment [91,92,93]. The KNN and RF classification algorithms [57] were applied to the 26 independent variables using the OTB 7.1.0 toolkit implemented in the QGIS.

2.4.1. K-Nearest Neighbor and Random Forest Classifiers

This study evaluated and compared the performance of two classifier algorithms: KNN and RF. For the KNN classifier, which has K values ranging from 1 to 35, tests were performed to identify the optimal k value for all training sample sets. The tests revealed an imbalance in accuracies, caused by low K values (overfitting) and very high K values (underfitting), which influenced the performance of the training data [48]. Considering the obtained results, the default value K = 32 defined by OTB was selected for providing a balance between classification accuracy and model generalization. This choice simplifies the process, reduces overfitting effects, balances bias and variance, and ensures classification stability [94,95]. This approach represents a viable alternative for users to initiate the use of the classifier without the need for manual adjustments of this parameter, enabling the achievement of satisfactory classification results [96].

The RF classifier is an algorithm that relies on the configuration of crucial parameters, such as the number of trees, the maximum depth of these trees, and the out-of-bag (OOB) error [97,98]. The optimization of these parameters significantly impacts the model’s performance and accuracy [30,41]. The validation dataset was used to generate an unbiased estimate of the error [99]. In this study, the RF model was configured with 500 trees for stability [34,100,101], a maximum depth of 25 to capture complex interactions [102], and an OOB error threshold of 0.01 as a rigorous stopping criterion [34], to maximize classification accuracy and minimize prediction error. This number of trees provides a good balance between processing time and accuracy. Multiple studies [51,88,95,103,104] suggest that increasing the number of trees beyond this point is unlikely to significantly reduce errors.

2.4.2. Accuracy Assessment

By comparing the classified map with the reference image, using the information from the validation samples, accuracy statistics can be derived by the confusion matrix. The confusion matrix is evaluating the degree of errors and classification accuracy in classed maps [105,106]. The confusion matrix assesses the degree of errors and classification accuracy in the classified maps. LULC map accuracy was tested for the overall (OA), producer’s (PA), and user’s (UA) accuracies and kappa statistics, using the validation dataset (30% of the validation samples) [107,108]. The AO, PA, and UA have been ranked as [101,102] very high (over 90%), high (80–89%), acceptable (70–79%), and low (less than 69%). The kappa coefficient (K) proposed by [109] is most frequently used to test map accuracy. This index is calculated based on the confusion matrix, incorporating both the hits and the omission and commission errors to provide an accurate assessment of agreement [14]. The kappa evaluates the degree of agreement among the classified map and ground truth [110,111]. For a thematic map K > 0.8, the agreement is high [109].

3. Results

3.1. Classification Accuracy

The KNN and RF classification achieved overall accuracy (OA) values greater than 75% for all regions. Both algorithms performed well in terms of classifying the data accurately. However, the RF algorithm stood out for producing OA and k values greater than KNN (approximately 3.5%). The RF algorithm achieved the highest OA value of 92.16% (K = 0.91) in the AC region, while the KNN algorithm had the second highest OA value of 88.69% (K = 0.87) in the same region (

Table 3. Overall accuracy (OA) and kappa coefficient (K) assessment of the RF and KNN algorithms.

NUTS III Regions	AC		AL		BA		AA		AG
Classifiers	KNN	RF	KNN	RF	KNN	RF	KNN	RF	KNN	RF
OA (%)	88.69	92.16	83.03	87.49	80.99	86.88	80.25	85.24	75.43	81.86
K	0.87	0.91	0.80	0.86	0.78	0.85	0.77	0.83	0.72	0.79

). On the other hand, the minimum values of OA and K were reached in the AG region for both algorithms. Among the NUTS III regions, the AC region had the best classification accuracy with RF and KNN, followed by the AL, BA, AA, and AG regions, in descending order.

3.2. Forest Classes Level Accuracy Assessment and Classified Maps

In all regions considered in this study, the classes with other uses, such as water surfaces and bare soil and artificial surfaces, reached values of UA (commission errors) between 70.81% and 99.92% and values of PA (errors of omission) between 70.11% and 99.54% in both tested classifiers. These results revealed that these sets had a good distinction between forest classes, that is, good spectral separability. Despite the good distinction of the spectral response, this study focuses on evaluating the performance of the classifiers in distinguishing the forest classes. Thus, a deeper evaluation was necessary to understand the contrasts between the forest classes and the reasons for classification errors.

In the forest classes set, the RF classification algorithm was the one that obtained the best PA and UA results in all regions studied (

Table 4. Producer (PA) and user (UA) accuracy assessment for the forest classes set obtained by KNN and RF classification.

NUTIII Regions		AC		AL		BA		AA		AG
Forest Classes	Accuracy	KNN	RF	KNN	RF	KNN	RF	KNN	RF	KNN	RF
Eucalyptus spp.	PA(%)	81.59	87.85	70.69	77.61	74.38	83.47	77.38	79.14	77.43	85.40
	UA (%)	80.03	88.17	68.91	77.90	73.05	84.25	77.66	80.30	69.27	78.72
Pinus pinea	PA(%)	82.91	89.06	67.44	76.13	80.00	85.33	66.59	81.10	64.96	72.45
	UA (%)	83.97	89.75	77.62	79.90	73.75	80.80	71.98	75.90	65.09	77.05
Pinus pinaster	PA(%)	-	-	74.48	77.80	-	-	75.25	81.15	54.09	67.24
	UA (%)	-	-	68.95	76.27	-	-	79.34	82.75	65.81	72.72
Quercusrotundifolia	PA(%)	82.45	85.79	-	-	64.09	73.61	82.06	84.55	71.63	78.58
	UA (%)	75.53	80.94	-	-	64.87	75.42	68.71	79.69	63.06	71.67
Quercus suber	PA(%)	84.10	86.77	78.31	84.52	69.40	78.22	63.19	74.45	61.28	72.04
	UA (%)	86.99	90.00	70.46	79.51	76.27	82.22	74.13	78.55	57.00	65.47

). The RF algorithm had a PA range from 65.24 to 89.06% and a UA range between 65.47 and 90.00%, while the kNN classifier had a PA range from 54.09 to 84.10% and a UA range from 57.00 to 86.99%. The region that presented the best results for PA and UA was the AC, using the RF classifier. The lowest individual accuracy values between the forest classes were found in the AG region, by both classifiers.

When verifying the accuracy assessment through visual inspection of the classified map and the reference data, a very similar spatial distribution of forest cover was identified in both maps. In AA and AC regions, the dominant forest cover consisted of Q. suber and Q. rotundifolia, as Montado systems. Additionally, small mosaics of Eucalyptus spp. and P. pinea were identified, as well as P. pinaster stands in the AA region, located within the Serra de São Mamede Natural Park (located in the east of this region (39°18′47″ N; 7°21′36″ W)). In BA region, the predominant cover was spatially dispersed Q. rotundifolia with a small concentration of P. pinea in the northeast of the region. In the AL region, a higher presence of P. pinaster was observed in the coastal region, along with mixed stands of Q. suber and P. pinea further north. In this region, to the west, there were Eucalyptus spp. and P. pinaster stands near the Serra de Monchique (located in the west part of this region (37°19′0″ N; 8°33′15″ W)), while a higher spatial distribution of Q. suber was found in the central part of the region. The resulting maps from the classification process using KNN and RF, depicting the main forest cover types and other classes, are shown in Figure 3.

In the AC and BA regions, P. pinaster occurs in very small patches, making its identification difficult. Therefore, this class was not considered for these two regions under study. The same happens for the species Q. rotundifolia in the AL region, where its presence is scarce and dispersed.

4. Discussion

To map the LULC in the NUT III regions, supervised learning algorithms (KNN and RF) were used with a greater focus on the spatial distribution of forest species in complex Mediterranean ecosystems. In the classification step, the same algorithm processing condition (with training and validation datasets) was used to compare the performance of the two algorithms. The accuracy in the production of thematic maps was evaluated by statistical measures obtained by the confusion matrix. The RF algorithm presented greater accuracy and kappa values when compared with the KNN algorithm results, in all regions studied. This was due to its ability to work with numerous input variables, known as big data [112].

Our classification results using the RF algorithm on Sentinel-2 imagery for LULC mapping in the Mediterranean region significantly outperformed those obtained in comparable studies [88,113,114]. For instance, in a classification study conducted at the Köyceğiz Forest Management Unit, in Turkey, species of Genista acanthoclada, Erica spp., Phillyrea latifolia, Quercus spp., Olea europea, and Arbutus andrachne were classified using Sentinel-2 images using the RF algorithm implemented in the R programing language. The overall accuracy was 78% and kappa 73%, and the accuracies (PA and UA) for each class were greater than 70% [113]. In forest areas located in the eastern part of the Tuscany region, Italy, Sentinel-2 data were evaluated to classify forest classes proposed by [115] in a Mediterranean environment. In the Tuscany region [114], when classifying European Forest Types (EFTs) using the RF classifier, they reached an OA value of 92.7% (k = 92.6%), and for all forest classes, the UA and PA values were greater than 83% using the R programing language. These results reached values close to those found in our work. The reason for these results was that the RF algorithm works better with numerous input variables and is less sensitive to imbalance in training and noise in the training set. Still, other reasons can be cited, for example, it is easy to parameterize, presents good ability to manage collinear resources [116], and is robust against overfitting and outliers [117]. In addition to classification studies in the Mediterranean ecosystem, other authors point to the performance of RF and KNN classifiers on single-date Sentinel-2 images, with good accuracy. Vasilakos et al. [4], using Sentinel-2 data to classify landscapes from the island of Lesbos, Greece, found that when testing six classification algorithms in MATLAB, the KNN and RF classifiers had the better overall accuracy of 89 and 90%, and a kappa coefficient of 86.36% and 87.89%, respectively. In addition, the forest classes evaluated as Oak sp., Pinus brutia, Castanea sp., and Pinus nigra did not show great variation in the accuracy values of users and producers [4]. This was an indication that both classifiers may be eligible to produce good forest cover maps.

In comparison with previous studies, our approach utilizing the OTB plugin achieved higher accuracy results, comparable to those developed using programming languages. When mapping LULC with KNN and RF algorithms, both demonstrated good performance, with the RF algorithm yielding higher results. The evaluation metrics of the models, using OTB classifiers, exhibited considerably higher values for overall accuracy and kappa coefficient, evidencing a substantial increase in classification accuracy. Additionally, the obtained producer’s accuracy (PA) and user’s accuracy (UA) were equivalent to or higher than those reported in the literature, indicating robust consistency across all mapped forest classes.

4.1. Contribution of Independent Variables to Gains in Accuracy

Although our study did not quantitatively assess the relative importance of the independent variables in the classification due to QGIS OTB limitations, the extensive literature review provides solid evidence on the crucial role of these variables in increasing the accuracy of LULC maps. In classification contexts, the exclusive use of spectral bands available in Sentinel-2 often proves insufficient for classifying dense and complex mosaics of different types of forest cover [118]. This is due to the presence of mixed spectral signals from vegetation and similar land covers [119]. Consequently, the incorporation of vegetation and texture indices emerges as a determining factor in accuracy gains in classified maps.

The literature has highlighted the importance of certain Sentinel-2 spectral bands for land use and land cover (LULC) classification, particularly in vegetation classification processes. Previous studies have identified the red, NIR (B8), and SWIR1 (B11) bands as the most significant for mapping different vegetation classes [57,120,121,122]. The red-edge and SWIR bands have proven crucial for mapping tree species and crop types [88], while the addition of spectral bands with 20 m spatial resolution, resampled to 10 m, allowed for a significant increase in overall accuracy (OA) of single-date classification [120].

The red-edge bands are sensitive to the variability of chlorophyll a and b [84,122,123], while NIR is sensitive to the variations within and between canopy radiation scattering [57], and SWIR to water content of forest species [124,125]. Other LULC studies have also identified these spectral bands as important variables for classification, especially for forest classes [88,92,126]. In Germany, with SA2S images from one date, the bands with the best predictive ability for forest species were red-edge 1 (B5), SWIR 1 (B11), and the blue band (B2) [88]. For forests in the Roorkee region, India, the NIR band (B8) of SA2S data from a single date was able to separate forest classes in a range of densities [92]. For Eucalyptus spp. species, the bands with the best performance in separating Eucalyptus spp. from other land use classes were the green (B3), red-edge 3 (B7) red-edge 4 bands, and/or narrow NIR (B8A) [126].

The inclusion of auxiliary data, such as VIs and textures, is a common approach to formulate a workflow that influences accuracy gains in classification. Our decision to consider certain VIs in this study was based on their demonstrated importance in LULC mapping in Mediterranean ecosystems [127]. From the traditional IVs, SAVI contributed to the soil adjustment constant in the background of the vegetation signal, an important factor when it comes to studies in the dry season and in semi-arid regions [80], and GNDVI, used to estimating photo synthetic activity, thus contributing to separating some forest species [128]. Moreover, EVI showed higher sensitivity in capturing the variability of the canopy in forest systems, due to the changes in namely leaf pigments (chlorophyll and leaf nitrogen) and variability in biomass [82]. The new VIs used in this study, depending on red-edge bands for their calculation, are characterized by lower absorption by chlorophyll, but are sensitive to changes in its content, reducing the saturation effect and enhancing the sensitivity of these VIs to moderate-to-high vegetation densities, reducing the traditional VI saturation effect [16,123]. Research conducted in the Extremadura region of Spain, characterized by Mediterranean vegetation, evaluated multiple spectral indices derived from Sentinel-2 imagery [73]. The study tested various indices, including Brightness Index (BI), Color Index (CI), Green Normalized Difference Vegetation Index (GNDVI), Second Modified Soil-Adjusted Vegetation Index (MSAVI2), Normalized Difference Index (NDI45), Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Soil-Adjusted Vegetation Index (SAVI). In the classification of forest classes using Random Forest algorithms, only CI showed significance. The forest categories included evergreen conifers (Pinus pinaster, Pinus pinea), dense evergreen broadleaf stands (Quercus rotundifolia, Quercus suber with canopy cover >60%), sparse evergreen broadleaf areas (20–60% canopy cover), and deciduous broadleaf forests (Quercus pyrenaica, Castanea sativa), along with nine additional land use classes [73]. Thus, these VIs can contribute to the spectral separability of forest species with proximal spectral response, such as Q. rotundifolia and Q. suber.

The incorporation of texture indices, specifically GLCM metrics, has also proven valuable in various LULC studies, demonstrating improvements in classification accuracy, as observed in our research [4,101,129,130]. In this study, GLCM proved to have higher sensitivity to the variation of crown cover and abrupt variation in digital numbers neighboring image features, which are predominant in the dry season [131]. Moreover, when studying forest classes with heterogeneous and complex spatial distribution, these bands highlight the contrast at stand boundaries [132,133]. Zheng et al. [68] found that incorporating texture measures into Sentinel-2 spectral data in the RF classification algorithm resulted in a significant increase in LULC accuracy, including in agricultural and vegetation classes. This approach proved effective in mitigating the limitations inherent to pixel-based map classification, addressing the insufficiencies in spatial detail and image sharpness of this algorithm [68].

In the Mediterranean region, Fragoso-Campón et al. [73], investigating the use of a single date for Sentinel data, found that Sentinel-2 texture metrics were very important for optimizing the classification algorithm, especially those derived from visible (B2 and B3), NIR (B5 and B6), and SWIR (B11) bands, which ensured good classification results in the Extremadura region, Spain. Blanco-Sacristán et al. [134] conducted a study using Sentinel-2 data to classify shrub populations in areas affected by historical fires in Sierra Nevada, Spain. In assessing the regeneration process and conservation status of vegetation, the incorporation of textural features derived from GLCM demonstrated a significant contribution to increasing mapping accuracy [134]. Morell-Monzó et al. [135], using very-high-resolution (VHR) images and GLCM textures in mapping abandoned citrus fruits in Oliva, eastern Spain, found that texture features with a 9 × 9 window size had the best overall accuracy with the Random Forest model.

Other studies conducted by Wang et al. [136] found that textures improved the accuracy of vegetation classification in wetland areas using high-resolution Pléiades satellite images. Feng et al. [137], investigating urban vegetation classification using the RF algorithm and ultra-high-resolution (UHR) digital images obtained by unmanned aerial vehicles (UAVs), highlighted the significant contribution of GLCM textural features in gaining accuracy and eliminating image blurriness, inconsistencies in vegetation classification, and especially the “salt and pepper” effects. Furthermore, the authors observed that the integration of textural metrics compensated for the spectral limitation’s characteristic of UAV images [137]. This study highlights the relevance of optical sensors, such as Sentinel-2, in the classification of tree species, based on evidence from the literature. Strategically, it incorporated a set of vegetation indices and textural features to enhance the spectral separability of forest classes and optimize overall classification accuracy. This approach aims to overcome the inherent complexities in classifying Mediterranean forest ecosystems characterized by complex mosaics. The proposed methodology seeks to optimize the application of remote sensing techniques in contexts of high landscape heterogeneity, where precise discrimination of forest classes represents a significant challenge.

Our approach of incorporating vegetation indices and texture, in addition to the standard Sentinel-2 spectral bands, aligns with these findings from the literature. The observed improvement in our classification accuracy suggests that the inclusion of these additional variables played a significant role in enhancing our results obtained from open-source software and plugins.

4.2. The Forest Occupancy Classification Challenges with Images from Sentinel-2

In Mediterranean ecosystems, LULC classification studies using Sentinel-2 images have demonstrated that spectral separability is challenging due to the spectral overlap among pairs of forest classes, montado systems, and shrublands, as observed in the autonomous region of Valencia, Spain [138], and among distinct types of broadleaf and coniferous forests in Italy [139]. When it comes to classifying a landscape dominated by fragmented forest stands, remote sensing-based landscape classification remains challenging, as most pixels exhibit similar spectral signatures influenced by ground reflectance or other species and shadows [140,141].

In our study, the classification reliability was verified through visual analysis, comparing the classified map with a reference map, such as high-resolution images from Google Earth. In general, the Q. rotundifolia class exhibited several pixels resembling the Q. suber class, shrub and agriculture surface. The mixed species composition in agro-silvopastoral environments and along stand boundaries resulted in categorization errors for Q. rotundifolia and Q. suber stands in montado regions. Particularly in montado systems, classifiers might struggle to detect the spatial arrangements of species due to significant variations in mosaic patterns and distribution of forest species [18,142,143,144]. The spectral similarity between the transition boundaries of regions with center-pivot agriculture, olive groves, and variously oriented vineyard planting lines across all analyzed locations also led to the confusion of some pixels in the Q. rotundifolia and Q. suber classes. In the AG and BA regions, some Q. rotundifolia stands, still young or spatially dispersed in the landscape, had some pixels confused with the shrub surface class. The understory vegetation or soils reflectance was reported in other studies in young stands, caused by canopy gaps [145,146]. In the AG region, signatures similar to Q. suber were identified within mosaic fragments featuring agricultural surfaces, along with plantations of other crops like Ficus carica, Prunus dulcis, Ceratonia siliqua, and Citrus spp.

In regions characterized by steep terrain, situated in AA (Natural Park of Serra de São Mamede), AL (Serra do Cercal), and AG (Serra do Caldeirão and Serra do Monchique), the P. pinaster class exhibited a higher classification error along boundaries of the land use patches with Eucalyptus spp. stands. In satellite optical imagery, remote sensing techniques do not account for topographic effects [147,148]. Therefore, our accuracy results might have been influenced by the presence of mountainous areas. Slopes that receive varying sunlight impact the proportion of light returning to the sensor due to changes in solar geometry and topography [149,150,151]. In these regions, we found greater challenges in defining Regions of Interest (ROIs) due to the vegetation’s heterogeneity, comprising other broadleaf species, shrubs, and young cork oaks. For these reasons, there was a heightened spectral confusion between the shrubland class and the Q. suber class, which is characteristic of mountainous regions, valley bottoms, north-facing slopes with less solar exposure, and coastal areas (AL and AG). Conversely, the gently undulating relief of the AC region may have contributed to higher accuracy in the LULC classes [152]. These justify the lower accuracy obtained for these regions.

Among the pairs of classes Q. suber and P. pinea, the confusion between them was nearly equal in the AL region. This is due to the more frequent occurrence of mixed stands of P. pinea with Q. suber in this region. However, across all studied regions, the classes Eucalyptus spp., P. pinea, and P. pinaster exhibited classification errors due to spectral confusion between riparian species and orchards with similar spectral signatures, as well as mixed stands. The P. pinea and P. pinaster stands were gradually replaced by Eucalyptus spp. due to decreased economic interest in resinous forests and a preference for rapidly growing and economically viable eucalyptus stands [153]. On the other hand, the emergence of mixed areas dominated by pine and eucalyptus reflected the lack of management in P. pinaster stands and agricultural abandonment [153]. For these reasons, between the classes of P. pinaster and Eucalyptus spp., the classifier struggled to delineate the composition of smaller P. pinaster mosaics accurately intertwined with larger Eucalyptus spp. mosaics. Furthermore, in our study, we identified small fragments of the second rotation of Eucalyptus spp., sparse P. pinaster stands, and young P. pinea plantations that influenced the spectral response due to spatially open canopies and higher reflectance from the ground, rocks, and/or shrub surfaces [154].

The monitoring of land use and land cover (LULC) dynamics is essential for landscape-scale conservation and management. While satellite-based remote sensing constitutes the most feasible approach for large spatial scales, heterogeneous landscapes can pose significant challenges for accurate land cover classification and change detection [155]. A primary obstacle encountered in this study was the classification of forest cover types across extensive territorial expanses, necessitating field expertise, in situ knowledge, visual interpretation of landscape features through ancillary data, and the utilization of high-performance computing resources for image processing. Nonetheless, this investigation addresses a critical gap in the existing literature regarding the large-scale application of algorithmic approaches to LULC analysis in southern Portugal.

Additionally, incorporating multitemporal S2 imagery and leveraging the synergy between Sentinel-1 data, as demonstrated in studies [1,4,156,157,158], can enhance the classification accuracies of forest occupation classes. Combining texture features extracted from Sentinel-1 (S1) data with Sentinel-2 (S2) multispectral bands enhances classification capabilities. While radar signals from S1 are sensitive to geometric properties (e.g., roughness, texture, and internal structure), optical reflectance from S2 is influenced by vegetation physiology [159,160]. To further improve the non-parametric models used in this research, we propose (1) incorporating additional forest classes, (2) integrating S1 and S2 data into time series, and (3) including new vegetation indices, texture metrics, and elevation features. These enhancements can address existing gaps and potentially increase model accuracy.

5. Conclusions

In this study, we propose an integrative approach to mapping the composition and distribution of the main forest species on a regional scale in southern Portugal, which involves the use of machine learning classification algorithms and free open-source tools. In environmental management and sustainable development, comprehensive mapping of forest land use and cover provides relevant information for developing sustainable forest management practices, climate change mitigation, managing natural resources, and protecting biodiversity and contributes significantly to improving the accuracy of forest biomass models.

The methodology was developed for mapping forest species using Sentinel-2 data in southern Portugal. The performance of two machine learning algorithms, K-Nearest Neighbor (KNN) and Random Forest (RF) from the OTB plugin, were analyzed to map the main forest species dominating in the five regions of Alentejo and Algarve, Portugal. Despite the complexity of forest species occupancy with high intraclass variability, accurate mapping using OTB plugin classification algorithms was achieved using single Sentinel-2A data.

The RF algorithm achieved the best accuracy results in all evaluated regions when compared to KNN. The best classification accuracies with RF and KNN algorithms were achieved in the AC region (RF: OA = 92.16%, k = 0.91; KNN: OA = 88.69, k = 0.87), followed by AL, BA, AA, and AG. Incorporating ancillary data, such as vegetation and texture indices, into the classification process contributes to the good results obtained. In addition, images acquired in the summer contributed to a good classification due to the lower occurrence of clouds and the distinctive spectral behavior of the understory, as well as between forest classes and other uses. Considering the scale and scope of this investigation, which encompasses the entire southern region of Portugal, the present study significantly expands existing knowledge about the area’s forest cover. While previous research has focused on analyses of larger scales [161] or more restricted areas [9], this work fills a crucial gap by providing a more detailed assessment of forest cover across the southern Portuguese region.

The methodology developed in this study offers an efficient and cost-effective approach for producing forest classification maps, utilizing intuitive interface tools like QGIS, SNAP, and the OTB plugin. Sentinel-2 data emerge as a viable and free alternative for large-scale forest classification. Its high spatial and temporal resolution competes with higher-resolution commercial satellite data. Image processing tools from non-commercial software packages facilitate forest species mapping in the Mediterranean ecosystem. These allow users without advanced programming knowledge to achieve good accuracy in their mappings.

Future work perspectives involve expanding the methodology to include classes with lower representation, encompassing the central–northern region of Portugal, and potentially other European regions. Integrating other data sources can help validate the studied model, including field data integration, and merging data from satellites, integrating LiDAR (Light Detection and Ranging) data, high-resolution images obtained from UAVs, and synthetic aperture radar (SAR) data from Sentinel-1. Specifically, integrating LiDAR data, it is worth mentioning that according to the Portuguese Directorate-General for Territory (DGT), the government intends to produce land use and cover maps, as well as hydrography, altimetry, and the Digital Terrain Model based on LiDAR coverage [162]. This project aims to democratize access to national cartographic resources, sharing open data information and integrating cartography resulting from cooperation between various levels of Public Administration.