®®®® SIIA Público

Título del libro:
Título del capítulo: Characterization of wildfire severity using deep learning and satellite time series analysis

Autores UNAM:
LILIA DE LOURDES MANZO DELGADO;
Autores externos:

Idioma:

Año de publicación:
2025
Palabras clave:

BFAST; Convolutional neural network; LANDSAT; Landtrendr; Planetscope; Post-fire; Post-fire recovery; Satellite time; Sentinel-2; Wildfire severity


Resumen:

© COPYRIGHT SPIE. Downloading of the abstract is permitted for personal use only.Fire severity is a key indicator for assessing post-fire ecological impacts, as it supports detailed damage evaluations and informs strategies to accelerate ecological recovery. This study characterizes the severity of two wildfires that occurred in 2019 in temperate forests of the Sierra Madre Occidental (Western Sierra Madre mountain range, SMO) in Mexico, using a U-Net architecture with bitemporal Sentinel-2 inputs and attention-residual blocks. Severity masks were generated using thresholds from the differenced Normalized Burn Ratio (dNBR) and the Relativized Burn Ratio (RBR) indices and were refined with PlanetScope composites and 83 point measurements from a composite field-based severity index. Despite limited data for the extreme class (Intersection over Union, IoU = 0.68), the model achieved high segmentation performance (mean IoU, mIoU = 0.83) across five classes and showed improved metrics when the number of classes was reduced (mIoU = 0.87 for four classes; mIoU = 0.92 for two classes). Post-fire dynamics were evaluated using the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Moisture Index (NDMI) time series (1993-2025) from Landsat Collection 2, applying the LandTrendr and BFAST-Monitor algorithms, which confirmed trajectories consistent with the predicted severity levels. Results revealed stable trajectories for low and moderate severity classes, while high and extreme severity classes exhibited breaks in 2019 and estimated recovery times ranging from 6.6 to 15.4 years. These findings highlight the value in the application of Convolutional Neural Networks (CNNs) and satellite time series for fire severity characterization.


Entidades citadas de la UNAM: