Combining LiDAR and hyperspectral data for aboveground biomass modeling in the Brazilian Amazon using different regression algorithms

Author(s):Show less -

de Almeida, Catherine Torres ^[1] ; Galvao, Lenio Soares ^[1] ; de Oliveira Cruz e Aragao, Luiz Eduardo ^{[2, 1]} ; Henry Balbaud Ometto, Jean Pierre ^[1] ; Jacon, Aline Daniele ^[1] ; de Souza Pereira, Francisca Rocha ^[1] ; Sato, Luciane Yumie ^[1] ; Lopes, Aline Pontes ^[1] ; Lima de Alencastro Graca, Paulo Mauricio ^[3] ; Silva, Camila Valeria de Jesus ^[4] ; Ferreira-Ferreira, Jefferson ^[5] ; Longo, Marcos ^[6]

Affiliation

^[1] Natl Inst Space Res INPE, Caixa Postal 515, BR-12227010 Sao Jose Dos Campos, SP - Brazil

^[2] Univ Exeter, Coll Life & Environm Sci, Exeter, Devon - England

^[3] Natl Inst Res Amazonia INPA, Caixa Postal 2223, BR-69080971 Manaus, Amazonas - Brazil

^[4] Lancaster Univ Bailrigg, Lancaster Environm Ctr, Lancaster LA1 4YW - England

^[5] Inst Desenvolvimento Sustentadvel Mamiraua, Caixa Postal 38, BR-69553225 Tefe, AM - Brazil

^[6] CALTECH, Jet Prop Lab, 4800 Oak Grove Dr, Pasadena, CA 91109 - USA

• Document type: Journal article
• Source: REMOTE SENSING OF ENVIRONMENT; v. 232, OCT 2019.
• Web of Science Citations: 1

Abstract

Accurate estimates of aboveground biomass (AGB) in tropical forests are critical for supporting strategies of ecosystem functioning conservation and climate change mitigation. However, such estimates at regional and local scales are still highly uncertain. Airborne Light Detection And Ranging (LiDAR) and Hyperspectral Imaging (HSI) can characterize the structural and functional diversity of forests with high accuracy at a sub-meter resolution, and potentially improve the AGB estimations. In this study, we compared the ability of different data sources (airborne LiDAR and HSI, and their combination) and regression methods (linear model - LM, linear model with ridge regularization - LMR, Support Vector Regression - SVR, Random Forest - RF, Stochastic Gradient Boosting - SGB, and Cubist - CB) to improve AGB predictions in the Brazilian Amazon. We used georeferenced inventory data from 132 sample plots to obtain a reference field AGB and calculated 333 metrics (45 from LiDAR and 288 from HSI) that could be used as predictors for statistical AGB models. We submitted the metrics to a correlation filtering followed by a feature selection procedure (recursive feature elimination) to optimize the performance of the models and to reduce their complexity. Results showed that both LiDAR and HSI data used alone provided relatively high accurate models if adequate metrics and algorithms are chosen (RMSE = 67.6 Mg.ha(-1) RMSE% = 36%, R-2 = 0.58, for the best LiDAR model; RMSE = 68.1 Mg.ha(-1) RMSE % = 36%, R-2 = 0.58, for the best HSI model). However, HSI-only models required more metrics (5-12) than LiDAR-only models (2-5). Models combining metrics from both datasets resulted in more accurate AGB estimates, regardless of the regression method (RMSE = 57.7 Mg.ha(-1) RMSE% = 31%, R-2 = 0.70, for the best model). The most important LiDAR metrics for estimating AGB were related to the upper canopy cover and tree height percentiles, while the most important HSI metrics were associated with the near infrared and shortwave infrared spectral regions, particularly the leaf/canopy water and lignin-cellulose absorption bands. Finally, an analysis of variance (ANOVA) showed that the remote sensing data source (LiDAR, HSI, or their combination) had a greater effect size than the regression algorithms. Thus, no single algorithm outperformed the others, although the LM method was less suitable when applied to the HSI and hybrid datasets. Results show that the synergistic use of LiDAR and hyperspectral data has great potential for improving the accuracy of the biomass estimates in the Brazilian Amazon. (AU)

• FAPESP's process: 16/21043-8 - Airborne LiDAR for quantifying changes in biomass stocks and structural dynamics in fire-damaged forests in Central Amazon
• Grantee: Aline Pontes Lopes
• Support Opportunities: Scholarships in Brazil - Doctorate