|
High-Precision Wildfire Risk Mapping Using an Ensemble Machine Learning Model: Performance Assessment of Sentinel‑2 and Landsat‑8 Data
|
Mohaddeseh Mesvari   , Reza Shah-Hosseini * |
| Department of Photogrammetry and Remote Sensing, Faculty of Geodesy and Geomatics Engineering, University of Tehran, Tehran, Iran. |
|
|
Abstract: (362 Views) |
Objective: Forest fires are among the most destructive natural hazards, causing irreversible ecological and economic consequences. Accurate fire risk mapping is therefore a critical tool for effective prevention and management. This study aimed to develop an integrated framework based on machine learning algorithms and to evaluate the effect of satellite data spatial resolution on the accuracy of forest fire susceptibility modeling.
Method: To achieve this objective, spectral and thermal data derived from Landsat 8 (30 m spatial resolution) and Sentinel-2 (10 m spatial resolution) sensors were combined with topographic and climatic variables. Five well-established machine learning models were initially trained, and their outputs were subsequently integrated through a hybrid ensemble model to predict fire-prone areas. This approach enabled the exploitation of the complementary strengths of individual machine learning algorithms, improved the modeling of complex nonlinear relationships within the input data, and enhanced overall model stability.
Results: Exploratory analyses revealed that the Normalized Difference Vegetation Index (NDVI), the Normalized Burn Ratio (NBR), elevation, and land surface temperature were the most influential variables in distinguishing fire-affected areas. Quantitative evaluation demonstrated that the hybrid ensemble model implemented using Sentinel-2 data outperformed all individual base models, achieving an overall accuracy of 94.52%, an F1-score of 87.76%, and an Intersection over Union (IoU) value of 80.12%. These findings indicate that integrating the capabilities of multiple machine learning algorithms can substantially improve predictive accuracy and model robustness compared with relying on a single model. Furthermore, a comparison of the two satellite sensors showed that the higher spatial resolution of Sentinel-2 significantly enhanced fire susceptibility estimation and produced maps with superior spatial detail and boundary delineation by reducing spatial averaging effects.
Conclusions: The results demonstrate that integrating hybrid ensemble models with high-resolution satellite imagery can provide an efficient and reliable framework for proactive natural resource management and forest fire risk mitigation. |
|
| Keywords: Forest Fire, Machine Learning, Hybrid Ensemble Model, Sentinel-2 & Landsat 8, Remote Sensing. |
|
|
Full-Text [PDF 1514 kb]
(81 Downloads)
|
Type of Study: Research |
Subject:
Special
Received: 2026/03/26 | Accepted: 2026/05/6 | Published: 2026/06/10
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
