INVESTIGADORES
RIVERA Luis Osvaldo
congresos y reuniones científicas
Título:
Modelling and mapping forest structure in Argentina using remote sensing predictors
Autor/es:
SILVEIRA, EDUARDA M.O.; RADELOFF VOLKER; MARTINUZZI SEBASTIƁN; MARTINEZ PASTUR GUILLERMO; POLITI NATALIA; LIZARRAGA LEONIDAS; RIVERA LUIS; ROSAS YAMINA; OLAH ASHLEY; BONO JULIETA; PIDGEON ANNA
Reunión:
Congreso; AGU Fall Meeting 2021; 2021
Resumen:
Accurate forest structure data are crucial for the developmentof sustainable forest management anddevelopment plans. However, mapping forest structural variables, such as heightand basal area, is a challenge, especially in highly heterogeneous areas. Remotesensing offers an efficient tool for modelling and mapping forest attributesover large areas. Images from Sentinel-1 and 2 capture structural and greennessinformation, and thus are expected to capture forests attribute variabilitywell. Our goal was to model and map tree height and basal area over the~392,484 km2 forest of Argentina, based on Sentinel-1 and 2 imagescaptured from 2015 to 2020, and on auxiliary information. We used random forestregression algorithm models, and our dependent variable data came from 3,788field plots that were part of Argentina?s recently conducted national forestinventory (2015-2020). Our predictors were derived from Sentinel-1 and 2 imagescaptured from 2015 to 2020, as well as auxiliary information.  For Sentinel-1 we used both VV and VHpolarizations and calculated 1st and 2nd order textures (mean,standard deviation, homogeneity, uniformity, correlation, and shade) within a3x3 moving window size. For Sentinel-2, we derived EVI (enhanced vegetationindex), calculated DHIs (dynamic habitat indices: cumulative, minimum andvariation) and EVI median to generate 1st and 2nd ordertextures. As auxiliary information we used latitude and longitude geographiccoordinates. We found that variables from Sentinel-1 and 2, and geographiccoordinates, predicted height (R2=0.7 and RMSE=27%) and basal area(R2=0.73 and RMSE=57%) well. Additionally, predictor variableimportance analysis showed that latitude, longitude, 1st ordertextures from VV polarization and from DHIs were key sources of information forlarge-area forest structure modelling. We generated detailed and spatiallyextensive maps of height and basal area at 10-m spatial resolution forArgentina?s forested area. These map products provide a strong basis for forestresources planning and forest species distribution modelling, as well as toolsfor conservation efforts across the country.