1.

INTRODUCTION

Land cover/use (LCLU) mapping provides essential information in climate change research[1]. Classifying vegetation characteristics is a crucial process for terrestrial carbon estimation, which facilitates climate change mitigation[2]. Vegetation maps are also used, in a plethora of applications in the field of natural resources management such as forest inventory, wildfire mapping, and water resources management[3].

Traditionally, LCLU mapping is based on field measurements, providing accurate results, although time-consuming and costly, especially in large and remote areas. Contrarily, remote sensing (RS) offers the opportunity for accurate and cost-effective LCLU mapping on different scales, utilizing airborne and spaceborne sensors [4]. With recent developments in RS systems, satellites can provide data at various spatial and temporal resolutions[5]. In the last decades, many studies on LULC mapping and monitoring have been carried out employing multispectral imagery from satellites, such as Landsat, Satellite for observation of Earth (SPOT), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Moderate Resolution Imaging Spectroradiometer (MODIS), and more[3], [4], [6], [7].

However, several authors have reported that medium to low-resolution observations, have a negative impact on the accuracy of the final product[8]–[11]. To overcome these limitations, machine learning (ML) approaches have been applied, instead of more traditional approaches (e.g. Maximum Likelihood and Minimum Distance) to classify remotely sensed images. Over a wide range of ML methods (e.g. Decision Tree, K-Nearest Neighbor, Artificial Neural Network, and XGBoost) Random Forest (RF) and Support Vector Machines (SVM), have gained a lot of attention in recent studies, due to their ability to provide reliable classification results, in different biomes and vegetation characteristics [12], [13].

With full constellation employed, Sentinel- 2A and Sentinel-2B satellites (launched on 23 June 2015 and 7 March 2017 respectively) provide high spatial (10m, 20m and 60m) and temporal resolution (approximately 5-day revisit cycle). In this study, we introduce a classification approach, based on machine learning techniques, for mapping mainly forest vegetation in three study areas in Greece. More specifically, we implement RF and SVM classification algorithms and compare their performance in mapping accurately vegetation species and cover, by employing seasonal Sentinel-2 multispectral imagery. The areas, selected for classification present different vegetation characteristics, including Greece’s most dominant vegetation species. Therefore, one of the goals of this research is to also provide a reliable mapping methodology that could be used at national level for supporting forest management and forest inventory planning.

2.

MATERIALS AND METHODS

2.1

Study area

The study was conducted in 2021 in the municipalities of Arta, Pella, and Korinthos (Fig.1) with approximate areas of 401km², 2506km², and 2297 km² respectively. The landscape of the selected sites exhibits a complex vegetation structure and consists of various species, with coniferous and broadleaves trees, shrubs, and grasslands covering the main bulk of the area. Moreover, the selected study sites include some of the most widespread forest species in Greece, such as Quercus coccifera, Pinus halepensis, Pinus brutia, and Abies cephalonica. The climate in all areas can be characterized as typical Mediterranean, with hot, dry summers and mild, rainy winters.

Figure 1.

Study areas in Greece; Municipalities of Korinthos (left), Pella (center), Arta (right)

3.

RESULTS AND DISCUSSION

Overall, the results of this study showcased that RF provided stronger classification capability (OA= 0.96, 0.90, and 0.95 with kappa=0.94, 0.85, and 0.93) than SVM (OA=0.89, 0.85, and 0.81 with kappa=0.84, 0.77, and 0.85) in all tested regions (table 2). The highest classification performance for both classifiers was observed in the Korinthos region. This can be attributed to the fact that Korinthos consists of a more homogenous landscape than the rest of the study areas. Also, the spectral discrimination of the vegetation types is more explicit in Korinthos than in Arta and Pella.

Table 2.

Classification results in the three study areas using RF and SVM.

To access the performance of the best model (RF) for each vegetation type, Producer’s accuracy (PA) and User’s accuracy (UA) metrics were investigated (table 3). More specifically, the classes that were better distinguished were Evergreen Broadleaves, Grasslands and Beech achieving PA and UA above 90%. However, it should be noted that Beech was not observed in two of the three study areas. In the case of Oaks, RF achieved good results in all tested sites (above 80%), while the lower PA value was observed in Arta (0.8). This is due to the misclassification of pixels between Oaks and Evergreen Broadleaves, which can be explained by the fact that both vegetation types are phenologically similar. Fir also provided good results (above 85% in both metrics) in Korinthos and Arta. A lower PA value (0.83) in Pella, indicates a minor underestimation of Fir, as some pixels were omitted to Beech and Pines classes. Regarding Pines, over 0.95 values were observed in Pella and Korinthos. On the other hand, Pines cannot be distinguished clearly from Fir and Evergreen Broadleaves, in Arta (PA=0.65 and UA=0.75). For Broadleaves, although, high accuracy metrics were obtained in Korinthos, lower PA (0.74) was reported in Pella. After a closer examination of the results, it is observed that Broadleaves were misclassified as Oaks. The reason behind the confusion between these classes is the same described for the misclassification between Oaks and Evergreen Broadleaves in Arta. Overall, all classes provided good results (PA and UA above 70%). Only one exception can be observed, for Conifers in Korinthos, where significant omission errors were reported (PA=0.55). This can be attributed to the fact that Conifers and Evergreen Broadleaves could not be discriminated properly, as some species (mainly shrubs) tend to have similar phenological characteristics. The vegetation maps resulted from RF implementation are illustrated in figure 2.

Figure 2.

Vegetation maps (2021) for Arta (left), Korinthos (center) and Pella (right).

Table 3.

User’s and producer’s accuracy for vegetation types in the three study areas using RF.

4.

CONCLUSIONS

In this study we investigated the potential of Sentinel-2 imagery in vegetation mapping, using two popular ML methods, namely RF and SVM. The proposed method relies upon multi-temporal imagery, in order to capture the spectral variations, occurring due to different phenological phases. To test the proposed methodology, we selected 3 study areas in Greece, based on their different vegetation characteristics. The results of the study demonstrated that the combination of Sentinel-2 imagery with both methods can provide reliable vegetation maps in Mediterranean ecosystems. Both classifiers achieved high classification accuracies, with Random Forest outperforming SVM in all of the study areas. Despite the fact that this research reached its goals, some limitations should be further investigated in future work. More specifically, the rather small sample size could be enhanced, in order to avoid uncertainty in the results. Furthermore, the extraction of additional phenological traits, potentially utilizing time-series imagery, could contribute to better discrimination of Oaks and Broadleaves. Finally, a more holistic approach will also encompass a more detailed validation of the results, using very high-resolution imagery and/or field data. Overall, our study provides a reliable framework that can be used in the future as the basis for accurately mapping vegetation patterns in Mediterranean ecosystems, at regional and national scale.