Integrating Deep Learning with eXplainable AI for Pixel-Based Analysis of Wildfire Severity with USGS FIREMON dNBRs in Turkey

Abstract

Wildfires are a major natural catastrophe that disrupts the normal cycle of ecosystems, causing forests to be destroyed. Annually, a substantial amount of forest is devastated by wildfires around the globe. Reliable and accurate data about the burnt areas is crucial for assessing the amount of wildfire damage. Utilizing remote sensing and advance deep learning techniques provides significant advantages in enhancing the dependability and effectiveness of detecting burnt areas. This study examines the wildfires that occurred place during July 2023 around the Aegean region of Turkey. Within the scope of the study, classification was carried out with pixel-based Convolutional Neural Network (CNN) using Sentinel-2A imagery before and after the wildfires. Prior to classifying the severity of the fire, the dNBR values were computed and four distinct degrees of intensity were identified using the thresholds established in USGS FIREMON. Prior to generating wildfire severity classes for the training data set, dNBR values were computed and 4 distinct intensity levels were identified using thresholds specified in USGS FIREMON. Spectral indices (Burned Area Index-BAI and normalized difference vegetation index-NDVI) of pre- and post-fire Sentinel-2A images were calculated and included in the data set during classification. As a result of the burn severity classification, the overall accuracy was determined as 90.24% and the kappa coefficient was 86.98%. The results demonstrate that the model attained exceptional accuracy rates across all test data. Furthermore, the SHAP methodology, which is a globally explainable artificial intelligence method, was employed to comprehend the decision-making processes of the trained deep learning model and assess the efficacy of each feature inside the model. The SHAP findings revealed that both post-BAI and pre-BAI variables significantly influenced the decision-making process of the model. In conclusion, this study proves how effective deep learning technique with XAI method are in accurately assessing fire damage. © 2025 Elsevier B.V., All rights reserved.