Selective Ensemble Machine Learning for Mapping Potential Invasion of Vachellia drepanolobium in Ethiopian Savanna Ecosystem Using Sentinel-2 and Environmental Variables
Bibliographic Data
| ID | 19586629 |
|---|---|
| Authors | Gamtesa Olika (0000-0003-0132-0498), Gamtesa Olika Bulto, Gizachew Kabite Wedajo (0000-0002-6332-1042), Tadesse Woldemariam Gole (0000-0002-7621-2873), Desalegn Chala (0000-0002-8045-6950), Dessalegn Obsi Gemeda (0000-0002-8635-260X) |
| Year | 2026 |
| Volume | 24 |
| Pages | 101218 |
| Publication date | 2026-03-01 |
| Open Access | Yes |
| Type | ARTICLE |
| Venue | Trees Forests and People (JOURNAL) |
| Journal identifiers | ISSN: 2666-7193 • E-ISSN: 2666-7193 |
| Publisher | Elsevier BV (PUBLISHER) |
| DOI | 10.1016/j.tfp.2026.101218 |
| OpenAlex | W7133231121 |
| Language | EN |
| References cited | 38 |
Invasive plant species, particularly Vachellia drepanolobium , pose serious threats to biodiversity and ecosystem services in arid and semi-arid rangelands. Borana National Park (BNP) and surrounding districts are increasingly affected by V. drepanolobium ; however, spatial modeling of this invasion remains limited. In this study, we evaluated the performance of six models: Random Forest (RF), Support Vector Machine (SVM), Gradient Boosted Machine (GBM), Multilayer Perceptron (MLP), Generalized Linear Model (GLM), and Maximum Entropy (MaxEnt), using the Area Under the Receiver Operating Characteristic Curve (AUC), True Skill Statistic (TSS), and Correlation Coefficient (COR). From 43 candidate predictors, 13 independent variables were retained following multicollinearity screening based on pairwise Pearson correlations (r ≤ 0.8) and Variance Inflation Factor (VIF < 10). Key predictors included elevation, Blue Band reflectance, Normalized Difference Moisture Index (NDMI), and Modified Chlorophyll Absorption Ratio Index (MCARI). Results showed that RF (AUC = 0.998; TSS = 0.980; COR = 0.939), MaxEnt (AUC = 0.991; TSS = 0.951; COR = 0.925), and GLM (AUC = 0.986; TSS = 0.923; COR = 0.915) were the top-performing models. These were integrated into a selective ensemble, which achieved superior predictive accuracy (AUC = 0.9982; TSS = 0.960; COR = 0.969) and produced spatially consistent invasion hotspots. The ensemble model estimated that V. drepanolobium has invaded 1,057.5 km2 (3.32%) of the broader study area and 316.35 km2 (4.69%) of BNP, primarily affecting three of the five management blocks. These findings provide the first spatially explicit prediction of V. drepanolobium invasion in Ethiopia, offering a scalable framework for monitoring invasion risk in savanna ecosystems globally
Biodiversity · Ecosystem · Ensemble learning · Invasive species · Random forest · Hydrological Forecasting Using AI · Remote Sensing in Agriculture · Species Distribution and Climate Change
| Highly cited | No |
|---|