Unitemporal approach to fire severity mapping using multispectral synthetic databases and Random Forests
Nenhuma Miniatura disponível
Data
2020-11-01
Orientador
Coorientador
Pós-graduação
Curso de graduação
Título da Revista
ISSN da Revista
Título de Volume
Editor
Tipo
Artigo
Direito de acesso
Resumo
Fire severity assessment is crucial for predicting ecosystem response and prioritizing post-fire forest management strategies. Although a variety of remote sensing approaches have been developed, more research is still needed to improve the accuracy and effectiveness of fire severity mapping. This study proposes a unitemporal simulation approach based on the generation of synthetic spectral databases from linear spectral mixing. To fully exploit the potential of these training databases, the Random Forest (RF) machine learning algorithm was applied to build a classifier and regression model. The predictive models parameterized with the synthetic datasets were applied in a case study, the Sierra de Luna wildfire in Spain. Single date Landsat-8 and Sentinel-2A imagery of the immediate post-fire environment were used to develop the validation spectral datasets and a Pléiades orthoimage, providing the ground truth data. The four defined severity categories – unburned (UB), partial canopy unburned (PCU), canopy scorched (CS), and canopy consumed (CC) – demonstrated high accuracy in the bootstrapped (about 95%) and real validation sets (about 90%), with a slightly better performance observed when the Sentinel-2A dataset was used. Abundance of four ground covers (green vegetation, non-photosynthetic vegetation, soil, and ash) was also quantified with moderate (~45% for NPV) or high accuracy (higher than 75% for the remaining covers). No specific pattern in the comparison of sensors was observed. Variable importance analysis highlighted the complementary behavior of the spectral bands, although the contrast between the near and shortwave infrared regions stood out above the rest. Comparison of procedures reinforced the usefulness of the approach, as RF image-derived models and the multiple endmember spectral unmixing technique (MESMA) showed lower accuracy. The capabilities for detailed mapping are reflected in the development of different types of cartography (classification maps and fraction cover maps). The approach holds great potential for fire severity assessment, and future research needs to extend the predictive modeling to other burned areas – also in different ecosystems – and analyze its competence and the possible adaptations needed.
Descrição
Palavras-chave
Idioma
Inglês
Como citar
Remote Sensing of Environment, v. 249.