Evaluation of the impact of wildfires and predicting vegetation cover in the sub-humid area using Google Earth engine and Stacking Ensemble Models

Nouadi, Kamel; Alshehri, Fahad; Marouf, Nadir

doi:.0..

DOI: https://doi.org/10.4491/eer.2024.732

Evaluation of the impact of wildfires and predicting vegetation cover in the sub-humid area using Google Earth engine and Stacking Ensemble Models

Kamel Nouadi¹^†, Fahad Alshehri², and Nadir Marouf¹

¹Natural Resources and Development of Sensitive Environments Laboratory, Functional Ecology and Environment Laboratory, Hydraulic Department, Larbi-Ben-M'hidi University, Oum-El-Bouaghi, 04000, Algeria.
²Earth Science Remote Sensing Research (ESRS) Chair, Geology and Geophysics Department, King Saud University- Riyadh, 11514, Saudi Arabia.

Corresponding Author: Kamel Nouadi ,Tel: +213665456960, Email: kamel.nouadi@univ-oeb.dz

Received: December 30, 2024; Accepted: March 17, 2025.

ABSTRACT

Using Google Earth Engine (GEE) to assess wildfire impacts, this study focused on a sub-humid area affected by 2021 wildfires. Sentinel-2 imagery (10 m) was used to compute NDVI and NBR+ indices, revealing nearly 50% vegetation loss, with weakened correlation post-fire. Accurate NDVI prediction is critical post-fires due to environmental changes. Two standalone machine learning models, Random Forest and XGBoost, along with a Stacking model combining their outputs, were employed to predict NDVI for 2030–2080. Historical data (2000–2023) for NDVI, climate variables (precipitation, LST) from CMIP6 model ACCESS-CM2, and future projections were utilized. Model evaluation showed close results, with MAE (0.0306–0.0310), RMSE (0.0410–0.0419), and R² (0.521-0.541). Predictions suggest slight NDVI variations over future decades, with stacking and Random Forest models providing stability, while XGBoost exhibited higher variability. Compared to current values (post-fires) NDVI projections indicate a long-term decline in vegetation, exacerbating aridity, reducing atmospheric humidity, decreasing rainfall, increasing runoff, heightening soil erosion, and lowering soil permeability. The study demonstrates the stacking model's strength in handling spatial data and improving prediction accuracy for regions prone to fires. It emphasizes the importance of early warnings for vegetation degradation, supporting vegetation management and environmental planning, especially during emergencies.