@InProceedings{AlmeidaGaArOmJaPeSa:2019:CoReTe,
author = "Almeida, Catherine Torres de and Galv{\~a}o, L{\^e}nio Soares
and Arag{\~a}o, Luiz Eduardo Oliveira e Cruz de and Ometto, Jean
Pierre Henry Balbaud and Jacon, Aline Daniele and Pereira,
Francisca Rocha de Souza and Sato, Luciane Yumie",
affiliation = "{Instituto Nacional de Pesquisas Espaciais (INPE)} and {Instituto
Nacional de Pesquisas Espaciais (INPE)} and {Instituto Nacional de
Pesquisas Espaciais (INPE)} and {Instituto Nacional de Pesquisas
Espaciais (INPE)} and {Instituto Nacional de Pesquisas Espaciais
(INPE)} and {Instituto Nacional de Pesquisas Espaciais (INPE)} and
{Instituto Nacional de Pesquisas Espaciais (INPE)}",
title = "Comparison of regression techniques for LiDAR-derived aboveground
biomass estimation in the Amazon",
year = "2019",
organization = "Congresso Mundial da IUFRO",
abstract = "Light Detection And Ranging (LiDAR) is an active remote sensor
that has been successfully applied for characterizing canopy
structure, especially to estimate aboveground biomass (AGB).
Parametric models, mainly the linear regression with stepwise
feature selection (LMstep), are the most common approaches used
for estimating AGB. However, non-parametric machine learning
techniques, such as Support Vector Regression (SVR), Stochastic
Gradient Boosting (SGB), and Random Forest (RF), can better
address complex relationships between biomass and remote sensing
variables. Therefore, it is desirable to assess the performance of
different regression strategies. This study aims to compare eight
regression techniques for LiDAR-based AGB estimation: LMstep,
Linear Models with Regularization (LMR), Partial Least Squares
(PLS), K-Nearest Neighbor (KNN), SVR, RF, SGB, and Cubist. For
this purpose, 34 LiDAR metrics were regressed against AGB from 147
inventory plots across the Brazilian Amazon Biome. Models
performance were evaluated by the average Root Mean Squared Error
(RMSE) and R2 from a 5-fold cross-validation strategy with 10
repetitions. The Kruskal-Wallis test was used to evaluate
statistical differences among models. Results showed that LMstep
presented the highest RMSE (68.85 Mg.ha-1) and lowest R2 (0.66),
while SVR had the lowest RMSE (65.23 Mg.ha-1) and highest R2
(0.69). However, the differences in performance of the models were
not statistically significant. Thus, we confirmed the results of
previous studies that showed that simple approaches, such as
linear regression models, performed just as well as advanced
machine learning methods for estimating AGB based on LiDAR data.",
conference-location = "Curitiba, PR",
conference-year = "29 set. - 05 out.",
language = "en",
urlaccessdate = "05 jun. 2024"
}