Aboveground biomass estimates over Brazilian savannas using hyperspectral metrics and machine learning models: experiences with Hyperion/EO-1

ABSTRACT

We investigated the potential of hyperspectral remote sensing to estimate aboveground biomass (AGB) over the Brazilian savannas (Cerrado), the second-largest source of carbon emissions in Brazil. For this purpose, a Hyperion/Earth Observing-1 (EO-1) image was collected in the dry season at the Ecological Station of Águas Emendadas (ESAE). In order to estimate the AGB, we evaluated the performance of five machine learning models (Classification and Regression Trees – CART; Cubist – CB, Partial Least Squares Regression – PLS; Random Forest – RF; and Support Vector Machine – SVM) and four sets of metrics (reflectance, narrowband vegetation indices – VIs; absorption band parameters; and the combination of these attributes). The lowest root mean square error (RMSE) was obtained for RF using VIs (29%) and a combination of metrics (28%). For VIs, RF differed from CUB, PLS and SVM at 5% significance level. From cross-validation results, the RMSE was 26.36% for grasslands, 35.04% for open savannas, and 24.85% for dense savannas. The RF model with VIs had the most stable predictive performance across the models, as indicated by small variations in RMSE from CART to SVM. The five most important ranked VIs in the RF model were the Normalized Difference Vegetation Index (NDVI), Pigment Specific Simple Ratio (PSSR), Enhanced Vegetation Index (EVI), Red Edge Normalized Difference Vegetation Index (RENDVI) and Structure Insensitive Pigment Index (SIPI). Most of their relationships with AGB were non-linear. The resultant AGB estimates showed consistent results with a vegetation cover map of the ESAE. Areas of the ESAE with AGB lower than 10 Mg.ha⁻¹ were coincident with the occurrence of grassland physiognomies (savanna grasslands and shrub savannas), while areas with AGB higher than 25 Mg.ha⁻¹ matched the occurrence of dense savanna physiognomies (woodland savanna and dense woodland savanna). Grassland areas showed larger values of coefficient of variation (CV) than areas of dense savannas. These first-hand results set a baseline of models and metrics for AGB modeling of savannas during the future transition from current sampling-type hyperspectral missions (< 10 km of swath) to large-coverage hyperspectral satellites (> 100 km of swath).

KEYWORDS:

• Hyperspectral remote sensing
• aboveground biomass (AGB)
• savannas
• Cerrado
• machine learning
• Hyperion/EO-1

Acknowledgements

The authors are grateful to the Instituto Brasília Ambiental (IBRAM) for the research authorization 06/2015 (project number 391.000.740/2015). We also thank field assistance provided by the Empresa Brasileira de Pesquisa Agropecuária (Embrapa – Cerrados), especially by José Ferreira Paixão. Lênio S. Galvão was supported by the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) (grant number 301486/2017-4), while Ricardo Dalagnol was supported by the Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) (grant number 2019/21662-8). The authors thanks the anonymous reviewers for the nice comments and suggestions.

Disclosure statement

The data that support the main conclusions of this study are available at https://figshare.com/s/e12fb88de75eac0e6f6b.

Additional information

Funding

This work was supported by the Conselho Nacional de Desenvolvimento Científico e Tecnológico [301486/2017-4]; Fundação de Amparo à Pesquisa do Estado de São Paulo [2019/21662-8].