Assessing GEDI-NASA system for forest fuels classification using machine learning techniques

Hoffrén, Raúl (Universidad de Zaragoza) ; Lamelas, María Teresa (Universidad de Zaragoza) ; de la Riva, Juan (Universidad de Zaragoza) ; Domingo, Darío ; Montealegre, Antonio Luis (Universidad de Zaragoza) ; García-Martín, Alberto (Universidad de Zaragoza) ; Revilla, Sergio

Resumen: Identification of forest fuels is a key step for forest fire prevention since they provide valuable information of fire behavior. This study assesses NASA’s Global Ecosystem Dynamics Investigation (GEDI) system to classify fuel types in Mediterranean environments according to the Prometheus model in a forested area of NE Spain. We used 59,554 GEDI footprints and extracted variables related to height metrics, canopy profile metrics, and aboveground biomass density estimates from products L2A, L2B, and L4A, respectively. Four quality filters were applied to discard high uncertainty data, reducing the initial footprints to 9,703. Spectral indices from Landsat-8 OLI scenes were created to test the effect of their integration with GEDI variables on fuel types estimation. Ground-truth data were comprised of Prometheus fuel types estimated in two previous studies. Only the types that matched in each GEDI footprint in both studies were used, resulting in a final sample of 1,112 footprints. Spearman’s correlation coefficient, Kruskal-Wallis and Dunn’s tests determined the variables to be included in the classification models: the relative height at the 85th percentile, the Plant Area Index, and the Aboveground Biomass Density from GEDI, and the brightness from Landsat-8 OLI. Best performances were achieved with Random Forest (RF) and Support Vector Machine with radial kernel (SVM-R), which were lower including only GEDI variables (accuracies: RF and SVM-R = 61.54 %) than integrating the brightness from Landsat-8 OLI (accuracies: RF = 83.71 %, SVM-R = 81.90 %). These results allow validating GEDI for fuel type classification of Prometheus model, constituting a promising information for forest management over large areas.

Idioma: Inglés

DOI: 10.1016/j.jag.2022.103175

Año: 2023

Publicado en: International Journal of Applied Earth Observation and Geoinformation 116 (2023), 103175 [10 pp.]
ISSN: 1569-8432

Factor impacto JCR: 7.6 (2023)

Categ. JCR: REMOTE SENSING rank: 6 / 62 = 0.097 (2023) - Q1 - T1

Factor impacto CITESCORE: 12.0 - Earth-Surface Processes (Q1) - Computers in Earth Sciences (Q1) - Global and Planetary Change (Q1) - Management, Monitoring, Policy and Law (Q1)

Factor impacto SCIMAGO: 2.108 - Computers in Earth Sciences (Q1) - Management, Monitoring, Policy and Law (Q1) - Global and Planetary Change (Q1) - Earth-Surface Processes (Q1)

Financiación: info:eu-repo/grantAgreement/ES/DGA/S51-20R
Financiación: info:eu-repo/grantAgreement/ES/MCIU/FPU18-05027
Financiación: info:eu-repo/grantAgreement/ES/NextGenerationEU/MS-240621
Financiación: info:eu-repo/grantAgreement/ES/UZ/CUD2020-07

Tipo y forma: Artículo (Versión definitiva)

Área (Departamento): Área Análisis Geográfico Regi. (Dpto. Geograf. Ordenac.Territ.)