Multi-sensor satellite data fusion and machine learning for Eucalyptus mapping in Meket district, Ethiopia

Abstract

Eucalyptus supports Ethiopia’s economy and its zero-carbon strategy, yet its rapid expansion in the highlands of Ethiopia creates ecological concerns. For better management, accurate mapping is needed, but it is challenged by cloud contamination, spectral mimicry, and the requirement for high-resolution commercial imagery. Therefore, this study used an integrated multi-sensor satellite data, such as Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 Multi-Spectral Imagery (MSI), for reliable and cost-effective mapping of Eucalyptus trees in Meket district, Ethiopia. Sentinel-2 data is endowed with multi-spectral bands that are sensitive to chlorophyll and leaf water content, whereas Sentinel-1 SAR also functions in all weather conditions and can capture moisture and structural information of trees. Eighteen features from spectral bands, radar backscatter, and vegetation indices were fused using a feature-level fusion strategy and classified using Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART). Results show that using freely available satellite data, RF achieved the highest performance, with an Overall Accuracy (OA) of 90% and a kappa coefficient of 0.81 in detecting Eucalyptus trees. SVM also achieved nearly the same performance with a 1% difference from RF. The study concludes that using publicly available Sentinel-1/2 fusion data with an appropriate classifier provides cost-effective, reliable, and accurate results for mapping of Eucalyptus, and supports Ethiopia’s zero-carbon strategy. It also helps policymakers and planners by providing geospatial technology-based land use planning for monitoring and sustainable management of Eucalyptus trees.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-58862-8.

Introduction

Recently, Eucalyptus trees have become the dominant plantation tree species in Ethiopia, which was introduced from Australia in the late 19th century¹. The Eucalyptus covers approximately 20 million hectares globally; Ethiopia holds 506,000 hectares, accounting for 90% of its planted woodlands^2,3. Economically, Eucalyptus is important for its uses as fuelwood, construction materials, and a source of money, supporting livelihoods and reducing pressure on natural forests^4,3,5. Ecologically, it contributes to reducing soil erosion and surface runoff, and helps land rehabilitation, but also causes challenges such as water and nutrient depletion, reduces soil fertility, and causes allelopathic effects^3,6. The controversy and paradox about the ecological and economic impact of the Eucalyptus tree is a continued discussion in the research community. For the sustainable management and conservation of natural and plantation forests (Eucalyptus tree), mitigating environmental impact and decision-making about the zero-carbon strategy of the country, accurate and updated tree-specific information, such as Eucalyptus tree mapping, plays a great role^7,8.

Recent advancements in earth observation technology, such as high-resolution sensors and machine learning, have enabled timely forest mapping, tree species detection, and mapping⁹. However, a single sensor imagery has a limitation in reflecting the complex ground objects due to different factors¹⁰. For instance, Landsat imagery (30 m) has spectral mimicry problems as compared with the Sentinel-2 imagery (10 m), which has better spatial resolution that can be used to detect specific Eucalyptus tree vegetation¹¹. In optical remote sensing, cloud cover/shadows are the factors that cause the loss of spectral information, particularly in rainy seasons. Although high-resolution satellite data such as WorldView-2 (8 MultiSpectral Imagery (MSI) bands with 1.8 m resolutions and a 0.5 m panchromatic band)¹², GeoFen-2 (1 m panchromatic imagery and 4 multispectral bands at 4 m resolutions)¹³, and RapidEye (5 bands with 5 m Spatial resolutions)¹⁴ are available commercially¹⁴. Despite their high resolutions, they are not viable for developing countries due to their inaccessibility and unaffordable cost¹⁵. stated that updated and accurate classification maps can be obtained without depending on expensive and commercial products such as high-resolution MSI or Light Detection and Ranging (LiDAR) data. This study was driven by the need for reliable mapping and identification of specific plant species, such as Eucalyptus, with minimal cost.

The remote sensing data fusion approach provides an alternative to multiple sensors and provides better results than a single sensor¹⁶ and resolves single sensor limitations. Fusing of multi-sensor data is an alternative to using costly data for low-budget planning and development¹⁷. The Sentinel-2 mission satellite provides multi-spectral imagery with high spatial, spectral, and temporal resolutions. But, the Sentinel-2 mission has a limitation of cloud contamination, mainly in rainy seasons, and does not include panchromatic bands and thermal infrared bands used for cloud masking. In contrast, Sentinel-1 SAR imagery is valuable in regions with cloud cover and limited optical data availability¹⁸. SAR is functioning in adverse atmospheric and ground conditions, for example, it passes cloud/shadow and captures the canopy structures of trees, and helps to characterize the structural attributes of observed features¹⁹. The SAR imagery can be operated in all-weather conditions by collecting information and improving the separability of nearly related land cover features^20–22. Both SAR and optical sensors respond to the Earth’s surface in different ways, making the observations made complementary to each other¹⁸, and this increases the mapping accuracy in land cover mapping²⁰.

Beyond the satellite imagery employed for mapping and change detection, the choice of classifier plays an important role in achieving high classification accuracy. ML algorithms have different advantages in handling multidimensional data (i.e., RF), non-linear relationships (i.e., SVM), and heterogeneous land cover classes (i.e., RF). Several researchers used single classifiers for Land Use and Land Cover (LULC) mapping, species detection and some of the studies compared ML classifier such as RF and K‑Nearest Neighbor (KNN)²³; RF, SVM and CART²⁴; Artificial Neural Network (ANN), RF and SVM²; RF, SVM, and Boosted Regression Tress (BRT)⁹; RF, SVM and Gradient Tree Boosting (GTB)¹⁵; RF and SVM²⁵; RF, SVM, GTB and CART using Sentinel-2 with axillary data²⁶. For this study, the three Google Earth Engine (GEE) integrated classifiers (RF, SVM, and CART) were used and compared to identify the most performing algorithm for mapping of Eucalyptus trees using integrated Sentinel-1 and Sentinel-2 data.

Meket district was selected as the study area because the Eucalyptus tree is expanding due to its fast-growing nature, low cost of establishment, easy adaptability, and being a source of livelihoods²⁷. The previous study using Landsat imagery revealed that the Eucalyptus tree expanded from 2003 (6022 ha) to 2010 (7838 ha) and then to 2019 (7930 ha)²⁷. The diverse topography, agroecology, and land cover nature of the study area make it ideal for evaluating the effectiveness of Sentinel-1 and Sentinel-2 imagery fusion with an ML classifier for Eucalyptus tree mapping.

A multiple study such as^{23,2,28,29,24,9,26,30,12,24,31}, was conducted for mapping and species detection of Eucalyptus plantation using single sensor optical data which is often affected by atmospheric factor (cloud cover). Other studies also used the SAR imagery and commercial sensor imagery for Eucalyptus mapping¹². Some of the researchers also used the fused data’s such as Sentinel-1/2, and LiDAR for forest species and type identification³¹, Sentinel-1 & 2 for commercial forest species discrimination^32,18, Unmanned Aerial Vehicle (UAV) and Sentinel-2 for stand density estimation³³, ALOS PALSAR-2, Landsat, and Sentinel-2 for Eucalyptus mapping³⁴, and WorldView-2 and Satellites Pour l’Observation de la Terre (SPOT-7) for Eucalyptus mapping²⁵.

To the best of our knowledge, limited research has been conducted to evaluate the effectiveness of Sentinel-1 and Sentinel-2 data fusion for mapping of Eucalyptus in Ethiopia. A systematic ML classifier comparison using Sentinel-1 and Sentinel-2 dataset fusion within the GEE environment has not been adequately examined. These study gaps are viable for a developing country like Ethiopia, where the commercial high-resolution data are unaffordable for routine monitoring. Accordingly, the primary objective of this study was (i) to assess the effectiveness of fused Sentinel-1 SAR and Sentinel-2 MSI imagery for mapping Eucalyptus trees in Meket districts, Ethiopia, and (ii) to identify the best-performing classification algorithm among RF, SVM, and CART. The main contribution of this study was to evaluate the Sentinel-1 and Sentinel-2 data fusion effectiveness for Eucalyptus tree mapping, identifying the best performing algorithm for multi-sensor data using GEE for scalable applications, and provide cost effective solution for mapping of Eucalyptus trees in a developing country where commercial sensor data are limited.

Methodology

Study area descriptions

Meket District is one of eleven districts in the North Wollo zone, Amhara Regional State, Ethiopia (Fig. 1). It is approximately 670 km from Addis Ababa, Ethiopia, and 222 km from Bahir Dar, which is the capital city of the regional state of Amhara³⁵. Geographically, it is located between 11°30′–12°10′ N and 38°30′–39°30′ E with elevation ranging from 1461 (northwest) to 3471 (southeast) m above sea level (Figs. 1 and 5). It has three agro-ecological zones: Kola (13–25%), Woina Dega (55–65%), and Dega (20–22%)^35–37, and is dominated by Cambisols³⁷. Meket district has 36 kebeles, which is the lower administrative level, and mixed farming is the main livelihood, which mostly depends on rain-fed^38,36, with mean annual rainfall from 600 to 1000 m²⁷. The vegetation comprises natural and planted trees and shrubs, such as Acacia spp. And Eucalyptus plantation³⁶. The recent projected population size of the districts is 292,509, where 147,052 males and 145,457 females³⁹. The district covers a total area of 1905.57 km² (190,557.45 hectares).

Fig. 1.

Map of study area. (a) Amhara and Ethiopia locations in Africa. (b) Meket district and North Wollo zone location in Ethiopia; (c) Meket district. The administrative boundary data source from CSA (2016). The OpenStreetMap base map adopted from NEXTGIS QuickMapService Version 1.3.2 plugin in QGIS software. (https://plugins.qgis.org/plugins/quick_map_services/). The Meket district Sentinel-2 Land cover-2024 downloaded from https://livingatlas.arcgis.com/landcoverexplorer. The map generated by the authors using QGIS software 3.34 version (https://qgis.org/ ).

Fig. 5.

Eucalyptus tree patches, road access, and elevation. The elevation data download from https://earthexplorer.usgs.gov/. The road data downloaded from https://openstreetmap.org/. The map generated by authors using QGIS software 3.37 version, https://qgis.org/ .

Remote sensing data acquisition

GEE is a powerful cloud-based planetary-scale environmental data analysis platform that offers a vast amount of satellite imagery and geospatial data (100 petabytes), including Sentinel, Landsat, and Moderate Resolution Imaging Spectroradiometer (MODIS), as well as powerful analysis and visualization tools with built-in code editors (JavaScript or Python)⁴⁰. This study utilized freely available Sentinel-1 SAR and Sentinel-2 MSI imagery, provided by the European Space Agency (ESA). The Sentinel-1 mission has four different satellites, such as Sentinel-1 A (2014-continuing), Sentinel-1B ( 2016-22), Sentinel-1 C (2024- continuing), and Sentinel-1D (2025-)^22,41.

Among ESA mission satellites, Sentinel-1 SAR imagery is operated in the C-band at 5.405 GHZ, which is used for mapping surface roughness, moisture content, and earth surface structures by recording the microwave signals reflected from objects interacting with the Earth’s surface^41,21. The spatial resolution of Sentinel-1 SAR imagery depends on the mode: Interferometric Wide (IW) (5*20 m), strip and wave mode (5*5 m), and extra-wide mode (20*40 m)^42,43. This study used Sentinel-1 Ground Range Detected (GRD) collection, which is available in GEE environments since 2014 at 10 m spatial resolutions and with single co-polarization (VV) and dual band cross-polarization (VH)^22,41,44.

Sentinel-2 MSI is an orthorectified, atmospherically corrected surface reflectance imagery, which is available in GEE starting from 2017 and provided by ESA under the Copernicus program. It serves as the bridge between the traditional medium resolution sensors and newer high-resolution commercial sensors³², which provide the 13 MSI bands at different resolutions: 10 m (four bands), 20 m (six bands), and 60 m(3 bands) with 14-bit radiometric depth and 290 km field of view^13,20,45. The two commonly used strategies for fusing SAR imagery and optical imagery are pixel-level and feature-level methods³². For this study, a feature-level fusion strategy (multi-source feature stacking) was employed. The features (Eighteen) include spectral and backscatter bands, as well as spectral and radar indices (Tables 1 and 2).

Table 1.

Satellite mission and characterizations.

Satellite mission	Band name and description	Pixel size (meters)	Wavelength, width, and unit
Sentinel-2	B2 (Blue)	10	496.6 nm (S2A)/492.1 nm (S2B)
	B3 (Green)	10	560 nm (S2A)/559nm (S2B)
	B4 (Red)	10	664.5 nm (S2A)/665nm (S2B)
	B8 (NIR)	10	835.1 nm (S2A)/833nm (S2B)
	B5 (Red Edge-1)	20	703.9 nm (S2A)/703.8 nm (S2B)
	B6 (Red Edge-2)	20	740.2 nm (S2A)/739.1 nm (S2B)
	B7 (Red Edge-3)	20	782.5 nm (S2A)/779.7 nm (S2B)
	B8A (Red Edge-4)	20	864.8 nm (S2A)/864nm (S2B)
	B11 (SWIR-1)	20	1613.7 nm (S2A)/1610.4 nm (S2B)
	B12 (SWIR-2)	20	2202.4 nm (S2A)/2185.7 nm (S2B)
Sentinel-1	VV (Vertical transmit/vertical receive)	10 (GEE)	250 km swath width, dB42
	VH (Vertical transmit/horizontal receive)	10 (GEE)	250 km swath width, dB42

S2A: Sentinel-2 A; S2B: Sentinel-2B.

Table 2.

Optical and radar indexes, equations, and citations.

Spectral index	Formula	References
Normalized Difference Vegetation Index (NDVI)		53
Normalized Difference Moisture Index (NDMI)		54,55
Red-edge NDVI (RE-NDVI)		56
Plant Senescence Reflectance Index (PSRI)		57
Radar Vegetation Index (RVI)		58
Backscatter/polarization ratio(R)		59

In addition to the SAR and optical imagery, auxiliary data such as digital elevation, Shuttle Radar Topography Mission (SRTM) 1 Arc-Second Global, downloaded from the United States Geological Survey (USGS) Earth Explorer website (https://earthexplorer.usgs.gov/) in GeoTIFF format, were used. The road network also downloads from crowdsourced geographic data such as the OpenStreetMap website (https://www.openstreetmap.org/ ) (.shp format). The pre-processing, such as reprojections (EPSG:32637 - WGS 84/UTM zone 37 N), clipping, pixel resolution matching using Bilinear resampling (i.e., 10 m), was conducted using QGIS 3.34 version software.

Image pre-processing and land cover classes

Eucalyptus trees can be harvested in 4–8 years due to their fast-growing nature and can provide clear observation (in remote sensing) from bare soil to forest within 1–2 years⁷. The study conducted to identify Eucalyptus trees among other forest types utilizes the winter (December-February) window, which enables better separability between the two land covers²⁹. For this study, the seasonal window (01/January to 30/June 2025) was used to capture the growing season, ensure clear observation (low cloud and shadow contamination), and effective discrimination of Eucalyptus trees (spectral and structural separability) from other land cover types.

The pre-processing steps for Sentinel-1 SAR data include radiometric correction (converting backscatter intensity to Sigma-Zero), speckle filtering, and range-doppler terrain corrections using Digital Elevation Model (DEM)^21,46. The Sentinel-1 SAR data available in GEE was used in its analysis-ready form (processed with the Sentinel-1 Toolbox), which involves thermal noise removal, radiometric calibration, and terrain correction (using SRTM 30-m DEM). Speckle noise in Sentinel-1 VV and VH imagery can be reduced using a 3 × 3 Lee adaptive filter while preserving edges and spatial details^47,48. In addition to these pre-processing methods, a multi-temporal composite (median composite)^49,50, image sub-setting (clipping), and radar and spectral feature derivations were performed with the GEE environment. Median composite was used to reduce variability and outliers, producing a stable and representative image for analysis. Therefore, for this study, in addition to GEE standard pre-processing approaches, Sentinel-1 imagery was employed for filtering, speckle filtering (Lee filtering), polarization processing, resampling and reprojection (to match with Sentinel-2).

All the Sentinel-2 (10 and 20 m) with less than 10% of cloud cover, and Sentinel-1(10 m) features were harmonized to a 10 m spatial resolution (using bilinear interpolation). Cloud and shadow contamination reduce the data quality by 50%⁵¹. Cloud and cirrus pixels were excluded from Sentinel-2 imagery employing the cloud mask algorithm⁵⁰, applying QA60 bands to remove cloud contamination and retain only clear observations for analysis⁵². After cloud masking, Sentinel-2 surface reflectance’s values were scaled to a 0–1 range by dividing by 10,000 and fixing outliers⁵². Then, using field-based and visual reference data, the overall land cover and use types are divided into two major classes: Eucalyptus and Non-targeted land covers. The two land cover classes can be defined as follows.

Eucalyptus tree (ET) Refers to the spatially connecting Eucalyptus cover, which exhibits a close and semi-closed canopy with uniform plantation structures.

Non-targeted land cover (Non-ET) includes the agricultural lands, natural vegetations, settlements, grazing lands, woody vegetations, shrubs, bareland, and other land covers that do not meet the criteria defined for Eucalyptus.

Optical and radar-derived vegetation indices

The use of optical and radar bands with the vegetation and radar indices improves the classification and discrimination among the different land cover types²⁰. Studies like²⁹ used the Enhanced vegetation index (EVI), Plant Senescence Reflectance Index (PSRI), Shortwave Infrared Water Stress Index (SIWSI), and MERIS Terrestrial Chlorophyll Index (MTCI) to distinguish Eucalyptus from forest types and provided the strongest spectral separability during the winter season. In this study, six vegetation indices were calculated and added to the composite images as additional feature bands (Table 2).

NDVI is widely used to indicate green biomass, while NDMI is applied due to its sensitivity to canopy water content. Red-edge NDVI is highly sensitive to chlorophyll content and leaf structural characteristics, thereby improving vegetation and species discrimination. Additionally, PSRI is sensitive to the carotenoid-to-chlorophyll ratio and is commonly used as an indicator of vegetation stress and senescence. The VV/VH ratio captures vegetation depolarization. It helps distinguish Eucalyptus canopies from vegetation, with a small epsilon(ε) added to avoid division by zero (See the data repository in the data availability section). RVI was used to capture vegetation biomass and canopy development for distinguishing Eucalyptus. The classification used an 18-variable feature stack combining Sentinel-2 spectral bands (10) and vegetation indices (4) with Sentinel-1backscatter (2), brightness ratio (1), and biomass indicators-RVI (1) to robustly capture the spectral and structural signature of Eucalyptus. Among 18 features, the most important features were also identified using the RF.

Training sample collection and sampling strategies

Training samples and independent validation points for Eucalyptus tree and Non-targeted land cover classes were collected using GEE tools based on Sentinel-2 MSI False colour composites (FCC) imagery for the study window, supported by high-resolution Google Earth imagery^28,7,60,61. Besides the imageries, the detailed local knowledge of the study area was used to guide the identification of plantation boundaries and species types. To obtain additional information and supplement the training samples collected, Google Earth imagery was employed⁶². A binary classification scheme (Eucalyptus labelled as Class-1) and Non-targeted land cover labelled as Class-2) was applied to distinguish Eucalyptus plantations from all other vegetated and non-vegetated land cover types.

Training samples were extracted using a stratified random sampling approach to ensure the representativeness of the sample points per classes (Eucalyptus and Non-targeted land cover). Totally, 339 sample points for the Eucalyptus tree and 444 sample points for Non-targeted land cover were extracted, which are spatially distributed across the study regions (Fig.S1). The sample was extracted using the Google Earth Engine sampling tools called sampleRegions({.})⁶³.

Sample splitting and spatial autocorrelations

The spatial data splitting approaches (i.e., spatial K-fold (leave-one-out cross-validation), object, block, cluster, and buffer-based partitioning) that can be used when reference sample sets have spatial dependence are discussed in⁶⁴. The feature collection was randomly split into a 70:30 ratio^65–67 used for training or model calibration (548) and testing/validation (235) subsets using a fixed random seed (42) for reproducibility. The spatial dependence between training and validation sets may cause an over-estimation of the model predictions (unrealistic) and bias because of spatial overfitting⁶⁴. Therefore, when collecting the sample points, the samples are collected across the study area (Fig.S1). But the spatial autocorrelation may exist between validation and training sets and is acknowledged as a limitation.

GEE embedded ML algorithms and its parameters tunings

According to⁴⁵, the tree species identifications depend on the type of classifier, imaging window season, training sample size, and spectral and spatial resolutions. In the GEE platform, the three classifiers: RF, SVM, and CART were used.

The RF algorithm (ee.Classifier.smileRandomForest()) was introduced by⁶⁸. It is non-parametric techniques that are versatile and capable of handling numeric and categorical data with minimal pre-processing of data. RF consists of several ensembles of decision trees using bootstrap sampling and random feature selections, which provides the resilience against overfitting, enhancing generalization and robustness against outliers^45,63. In RF, based on the inputs and training datasets, the final predictions are determined by majority voting to improve accuracy and reduce overfitting⁴³. For each model, the parameter tuning was implemented distinctly to improve accuracy. After the multiple trials and errors, the optimal parameter of RF was configured with 150 trees, a minimal leaf size of 5, and a bag fraction of 0.5, with a fixed random seed (42) to achieve reliable classifications. This large number of trees (i.e., 150) helps to improve the correctness of classification and reduce overfitting by taking the mean of the predictions from the number of trees.

The SVM (ee.Classifier.libsvm()) classifies data by constructing a separating hyperplane/ multiple hyperplanes in a high-dimensional space, with support vectors near the margin determining its positions⁴⁵. The SVM algorithm identifies an optimal separating hyperplane between different classes in the feature space⁴³. After testing different combinations of parameters, the SVM model architecture was adopted using Radial Basis function (RBF) as the kernel type, the gamma parameter was set as 0.50, and the cost parameter was determined as 10. Among the different kernel functions (linear, polynomial, and radial basis function (RBF) and sigmoid), the RBF and Polynomial are most widely used⁶⁹. SVM with Radial Basis function (RBF) Kernel, effective for complex, non-linear classification problems.

CART (ee.Classifier.smileCart()) was developed and initiated by⁷⁰. It is the simplest decision tree node and is highly sensitive to the training data. Despite its dependence on training data, it remains a simple and robust model with proven durability over time⁷¹. For this study, the CART classifier was implemented using default GEE parameters, providing an efficient and interpretable baseline for distinguishing Eucalyptus from Non-targeted land cover.

Model validation and accuracy assessments

The model validation was performed using the test datasets, which used 30% sample collections (102 for Eucalyptus Tree (ET) and 133 for Non-targeted land cover) Non-ET)). For this study, among the different cross-validations, such as k-fold and others, the hold-out validation approach (70:30 split) was applied⁷². A confusion matrix was used to evaluate the model’s accuracy, including specificity, sensitivity, and producers’ and users’ accuracy². For this study, conventional accuracy assessment methods - such as confusion matrix, Kappa coefficient, overall accuracy, producer’s accuracy(recall), and user’s accuracy (Precision) - were employed. Overall accuracy (OA) is a basic image classification metric that measures the percentage of correctly classified instances (True positives (TP) and true negatives (TN) out of the total samples.

1

where FP is False Positive, and FN is False Negative. To examine the uncertainty of classification accuracy, for each OA metric, the 95% of Confidence Interval (CI) was considered. The kappa coefficient (K)-Cohen’s kappa- is a statistical measure that evaluates the level of agreement between observed classification results and those expected by chance in image classification.

2

where: Po refers to the observed agreements, calculated by proportion of corrected classification (sum of the confusion diagonal) to total observations, while Pe is the expected agreement by chance, derived from the confusion matrix marginal totals.

User Accuracy (UA) (precision)- is the ratio of correctly classified pixels to the total pixels assigned to a class.

3

Producer’s Accuracy (PA)(recall) is determined by the ratio of correctly classified pixels to all reference pixels in the class.

4

Post-classification spatial refinement and area analysis

Based on the best-performing model (RF or SVM, or CART), morphological filtering was applied to remove isolated misclassified pixels and improve spatial coherence in the Eucalyptus map. A focal minimum (2-pixel) erosion followed by a focal maximum (4-pixel) dilation reduces noise while preserving and refining true Eucalyptus path boundaries. In the Amhara region, Ethiopia, rural landholdings range from a minimum of 0.25 hectares (rain-fed),0.06 hectares (irrigated land) to a maximum of 7 hectares in highland/semi-highland and 10 hectares in lowland areas⁷³. Based on the minimum and maximum land size holding in the study area, the size filtering was applied to remove patches smaller than 600 m² and larger than 1*10⁵ m² (10 ha), retaining only realistic Eucalyptus plantation areas based on regional knowledge. For the remaining polygons, the geodesic area was calculated and converted to hectares to estimate the total and individual Eucalyptus tree extent. The mean patch size (ratio of total Eucalyptus area by number of patches) and patch density (i.e., number of patches divided by total area of region/study) were determined. All processes, including pre/processing, classification, and validation of the three models using both optical and radar remote sensing data, were conducted using the GEE (Fig. 2).

Fig. 2.

Study diagram (colour shows related themes). The authors develop the diagram using https://app.diagrams.net/.

Study flow diagram

See Fig. 2.

Results and discussion

Performance measurements and post-classification analysis

Performance analysis

The Google Earth Engine integrated machine learning classifiers (RF, CART, and SVM) classification accuracy and maps are presented in Table 3; Fig. 3, respectively. Using the fused Sentinel-2 MSI and Sentinel-1 Radar imagery (18 features) for mapping of Eucalyptus tree, the RF and SVM demonstrated better and closer performance than CART in classification accuracy, which is also supported by^69,61. Random Forest (RF) achieved slightly better performance, OA of 90% (95% CI 93–97%) and a kappa coefficient of 0.81 than SVM. This performance is due to the capability of the model in handling the overfitting, noise, and managing high-dimensional data. The feature-level fusion of SAR data (Sentinel-1) and Sentinel-2 MSI leads to the building of high-dimensional data (18 features) that can be handled using the RF algorithm. RF is also robust for outlier data and faster than other models like SVM. RF has the capability to manage multi-dimensional data (temporal and source), noise, outliers, and select the important variables. Compared with SVM, RF is also less sensitive to parameters and sample variabilities, and capable of distinguishing the different land cover classes at a large scale and capturing non-linear patterns.

Table 3.

Accuracy parameters by the three classifiers (RF, CART, and SVM).

Accuracy parameters	RF		CART		SVM
	ET	Non-ET	ET	Non-ET	ET	Non-ET
PA (%)	88	93	77	83	84	93
UA (%)	91	90	80	82	91	88
Kappa	0.81		0.61		0.78
OA (%)	90		81		89

PA stands for producer accuracy, UA stands for user accuracy, OA stands for overall accuracy, ET stands for Eucalyptus Tree, Non-ET stands for Non-targeted land cover.

Fig. 3.

Map of Eucalyptus and non-Eucalyptus areas (based on CART, SVM, and RF algorithms). The map generated by authors using QGIS software 3.34 version (https://qgis.org/).

Similar studies in Spain (using Sentinel-2 imagery, LiDAR, and topographic data)³¹, China (i.e., Sentinel data with other ancillary data)²⁶ also confirm that RF performs better than other classifiers. The study in Akaki river catchments, Ethiopia, by² to detect Eucalyptus woodlots indicated that the RF (OA = 96.3% and Kappa = 0.93) performed better than ANN (OA = 88.7%) and SVM (OA = 81.7%). The study conducted in Portugal for forest species mapping also confirms that RF (OA = 92.16%, Kappa = 0.91) outperformed KNN²³. The study conducted in South China for classifying Eucalyptus, Cunninghamia lanceolata, and Pinus Massoniana forest using Sentinel-2 bands, Spectral indices, and textures confirms that the RF performed well with an OA of 93.22% and a kappa of 0.90³³.

SVM was performed with an OA of 0.89 (95% CI 86–92%) and a kappa of 0.78 following the random forest classifier (Table 3). The adopted hyperparameter tunings (such as Kernel type = RBF, gamma = 0.50, and cost = 10) show their importance for the valuable and highest performance of SVM, and relatively low gamma and cost values also help to reduce overfitting. The study conducted in Colombia for LULC mapping by integrating radar and optical datasets showed that SVM achieved the best performance (OA = 88.75% and Kappa = 0.86) compared to RF and KNN²⁰. Another study conducted in Gonder, Ethiopia, using Sentinel-2 and Sentinel-1 data in GEE showed that SVM achieved better performance (OA = 0.91 and kappa = 0.80) than RF (OA = 0.81 and kappa = 0.81)⁴³. A study conducted to detect native and exotic plant species, using fused Sentinel-1 and Sentinel-2 imagery, arranged the performance of classifiers as follows: CART (OA = 91.3, Kappa = 0.896); RF (OA = 90.0, Kappa = 0.88), and SVM (OA = 82.0, Kappa = 0.78)²⁴. The study conducted in Southern Africa by⁷⁴ using Sentinel-2 MSI also indicates that Eucalyptus woodlots were successfully detected from other land cover using SVM with an OA of 85.07%. SVM is more appropriate for small datasets due to its complex training mechanism (which depends on the size of the datasets) and non-linear relationships. Therefore, it is not suitable for large datasets and does not proved feature importance⁶¹.

In contrast, the CART algorithm was the least performing classifier with 81% (95% CI 77–85%) as compared with RF and SVM (Table 3). The CART classifier has some advantages in managing both continuous and categorical variables; it is easily interpretable and visualizable, and handles the missing data (i.e., surrogating split)⁶⁹. Besides its simplicity and robustness, the CART model is highly dependent on the training datasets, and changes in training datasets also cause different model outcomes⁶¹. The disadvantage of both CART and SVM is misclassification between LULC classes.

The study conducted in South Africa by³² also confirmed that the OA increased to 87% and 88% (Kappa = 0.83–0.85) when fused Sentinel-1 radar imagery with Sentinel-2 MSI using Linear Discriminant Analysis (LDA). The study conducted in Johannesburg for mapping Eucalyptus trees using WorldView-2 and SPOT-2 data, RF (OA = 81.67% and Kappa = 0.78 for WorldView-2, and OA = 72.78% and Kappa = 0.67 for SPOT-2) outperformed SVM²⁵. Another study conducted in the upper Blue Nile basin, Ethiopia, using RF, SVM, and BRT classifiers, Sentinel-2, and environmental data for modelling Eucalyptus species distributions, shows that RF performed better than the others⁹.

Features importance

Among the 18 features, the most important in the RF classification for mapping Eucalyptus trees were B3 (Green), RE-NDVI, B11 (SWIR-1), NDVI, B2 (Blue), B12 (SWIR-2), PSRI, and B8A (Red-Edge-4) (Fig. 4). The least-performing features are RVI, VV/VH ratio, VH, B6, and VV (Fig. 4). The study for detecting Eucalyptus species using RF and WorldView-3 indicated that the coastal Blue band (B01), Blue band (B02), and the red-related bands (Red edge, NIR-2, Red, and NIR-1) were the most important variables¹². A study conducted for forest mapping using RF indicated that Elevation, B9 (Sentinel-2) slope, GLI (Green Leaf Index), VH (Sentinel-1), and aspect are the top-performing features⁶⁹. B3 (green), RE-NDVI, and B8A (Red-Edge) bands are mostly sensitive to chlorophyll content, whereas RE-NDVI, B8A (Red-Edge), and NDVI are sensitive to vegetation stress. Red-edge bands are an ideal spectral range used for species identification. B11 (SWIR 1) and B12 (SWIR 2) are also useful for detecting the moisture content of leaves, which helps to detect the Eucalyptus (due to high-water content) from other species.

Fig. 4.

Random forest feature importance.

Post-classification analysis

Currently, the uncontrolled expansion of Eucalyptus is a common practice in Ethiopia. Based on the minimum land size (0.06 hectares) and the best performing classifier (i.e., RF), a total of 4174 Eucalyptus patches were developed (Fig. 5). About 5,378.56 hectares of land are also covered with Eucalyptus trees. This large number of Eucalyptus patches indicates fragmented distributions and smallholder-based planting (woodlots) owned by individual farmers. In the western highlands of Ethiopia, in the expanse of productive lands, the Eucalyptus plantation expands from 908.87 ha (1991) to 26,261.9 ha (2021)⁷⁵. A similar study in North-western Ethiopia confirms that the Eucalyptus tree expands by 55% (1992 to 2021) and then 69% by 2021¹¹. The study conducted in the Gilgel Abay watershed also confirms the expansion of Eucalyptus trees⁷⁶. The UAV-LiDAR-based individual tree detection study in an Eucalyptus plantation indicated that, given a 2 m tree spacing and 4 m row spacing, the expected density is approximately 1250 trees/hectare⁷⁷.

The average patch size was 1.29 hectares, indicating that the Eucalyptus tree plantations are mainly small-scale. The patch density is also 0.0219 patches /hectares (2.19 patches per 100 hectares), indicating the widespread and fragmented plantation across the study area. As Fig. 4 shows, Eucalyptus trees are mostly grown along road corridors (Woldia-Gashena-Woreta roads) to enable easy transport of fuel wood and construction poles to consumers at high elevations, showing Eucalyptus’ role as a market-oriented livelihood strategy. This spatial expansion into ecologically sensitive highland areas raises concerns about ecological impacts and food security issues. Highland areas are densely populated, and land fragmentation is also relatively high, pushing households to plant Eucalyptus as an additional source of livelihood, fuel wood, and other uses.

Conclusion and limitations

Conclusion

This study demonstrated the effectiveness of fusing Sentinel-1 SAR and Sentinel-2 MSI imagery for Eucalyptus tree mapping in the Meket district of Ethiopia using GEE. It confirms that using freely accessible Geospatial Technology (GST) data enables accurate and cost-effective mapping of Eucalyptus trees where access to commercial high-resolution data is limited. The results demonstrated that RF achieved the performance, with an OA of 90% and a kappa coefficient of 0.81, followed closely by SVM (OA = 89%, Kappa = 0.78) and CART (OA = 81%, Kappa = 0.61). Without a statistical significance test (such as McNemar’s test), RF and SVM have comparable performance for mapping of the Eucalyptus tree mapping with little advantage of RF.

The RF performance is mostly related to its robustness and capability of managing multi-dimensional data (fusing multi-sensor data), and the non-linearity of the features. The synergy between Sentinel-1 SAR data and Sentinel-2 optical imagery proved critical in improving the separability of Eucalyptus plantations from other land cover types, especially under conditions of cloud contamination. The most important features of RF classification were B3 (Green), RE-NDVI, B11 (SWIR-1), NDVI, B2 (Blue), B12 (SWIR-2), PSRI, and B8A (Red-Edge-4), accordingly. From the total land of the study area, 5,378.56 hectares of land were covered with Eucalyptus trees, and patches are highly fragmented (i.e., 4174 Eucalyptus patches and an average patch area of 1.29 hectares) and mostly concentrated in near road corridors and highland areas.

For operational mapping of Eucalyptus trees in Ethiopian highlands, the results indicated that the fusion of SAR imagery and Sentinel − 2 MSI with a proper ML algorithm, such as RF, provides cost-effective, accurate, and reliable results, which help the decision makers, practitioners, and regional development agents to monitor the expansion of exotic species such as Eucalyptus tree and sustainable ecosystem management. Such approaches also help the stakeholders and decision-makers by supporting Sustainable Land Management (SLM), informing Ethiopia’s Zero-carbon strategy, and facilitating policy decisions on balancing the ecological impacts and economic benefits of Eucalyptus.

Limitations

Some of the limitation that acknowledge in this study are variation in Sentinel-2 spatial resolution (10, 20) specially in heterogeneous areas, cloud contamination, and computational constraints of GEE (while using for large area and high-resolution data) can cause classification uncertainty. In addition, the performance of the classifiers also varies with parameter tuning (determined by trial and error), size, and complexity of the study area. The spatial autocorrelation between the training and validation tests may exist, and spatial cross-validation was not conducted, which may limit the robustness of the accuracy assessments. To overcome the above limitations and advance mapping of Eucalyptus tree, this research could further enhance by incorporating spatial cross-validations to handle the spatial autocorrelations, deep learning algorithms (such as CNN) with comprehensive field data (such as stem density, basal area, canopy heights (using UAV or LiDAR ) to further refine the species identification, and apply for large scale forest monitoring leveraging GEE platform.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Setiye Abebaw Tefera (SAT) : Conceptualization, Methodology, Analysis, Writing-original draft, Writing-review and editing, Investigation, Software. Vijaya Lakshmi Thatiparthi (VLT) : Conceptualization, supervision, validation, writing review, and editing.

Funding

The authors declare that this study received no funding, grants, or financial support.

Data availability

The datasets are available in a GitHub repository (https://github.com/Setiye11/Eucalyptus-Mapping_Multi-Sensor-Data-ML.git), and ancillary data can be obtained from the corresponding author upon request (mailto: setiyeabebaw2005@gmail.com (SAT)).

Declarations

Competing interests

The authors declare no competing interests.