Predição de trocas de carbono entre a biosfera e a atmosfera na FLONA-Tapajós a partir de variáveis ambientais
Data
2022-11-25
Autores
Bauer, Lucas de Oliveira [UNIFESP]
Orientadores
Rizzo, Luciana Varanda
Tipo
Dissertação de mestrado
Título da Revista
ISSN da Revista
Título de Volume
Resumo
A floresta Amazônica desempenha um papel importante no balanço de carbono terrestre, atuando como um sumidouro de carbono através da atividade fotossintética, e, ao mesmo tempo, como fonte de carbono por meio das emissões por queimadas, áreas alagadas e processos metabólicos terrestres. As trocas de CO2 entre a floresta e a atmosfera podem ser estimadas a partir de observações diretas na superfície, pela técnica de covariância de vórtices turbulentos. Porém, tais observações possuem uma representatividade espacial pequena, que não pode ser extrapolada para toda a Amazônia devido à heterogeneidade do balanço de carbono na floresta. O uso de estratégias de Ciência dos Dados pode ser uma alternativa para ampliar a escala espacial das estimativas de balanço de carbono, desde que sejam conhecidas as relações entre os fluxos de CO2 e variáveis ambientais, que muitas vezes são relações não-lineares. Este trabalho tem como objetivo construir modelos de aprendizagem de máquina para prever as seguintes métricas de balanço de CO2: troca líquida de CO2 (NEE entre a floresta e a atmosfera), produtividade primária bruta (GPP) e respiração (Re). Para isso, foram utilizados dados diários de fluxos turbulentos e de variáveis ambientais monitoradas entre 2002 e 2005 na Floresta Nacional dos Tapajós (FLONA-Tapajós), na Amazônia. Como preditores, foram consideradas variáveis meteorológicas de superfície, fluxos de calor sensível e latente, espessura óptica de aerossóis e índice de área foliar. Foram desenvolvidos modelos de regressão por Random Forest (RF) e Redes Neurais Artificiais (RNA). Também foram construídos modelos para a classificação de cenários de fonte de carbono, sumidouro e condição neutra. Os modelos de regressão tiveram coeficientes de determinação (R2) entre 0,33 e 0,65 para os modelos de RF, e entre 0,44 e 0,58 para os modelos de RNA. O modelo de regressão de NEE por RNA reproduziu corretamente o comportamento sazonal e os valores extremos. Apesar da variável GPP ter alcançado os maiores valores de R2, ambos modelos de RF e RNA falharam na previsão dos valores extremos dessa variável. A acurácia dos modelos de classificação variou entre 61% e 70%, sendo que o método de RF apresentou melhor desempenho. Dentre as variáveis preditoras, aquelas que apresentaram maior relevância nos modelos construídos incluem: radiação incidente no topo da atmosfera, fluxos de calor, índice de área foliar e temperatura. Os resultados obtidos sugerem a viabilidade de predição de fluxos de carbono na Amazônia a partir de variáveis ambientais, constituindo o primeiro passo para a extrapolação de observações de fluxo locais para a escala regional. Já os modelos de classificação permitiram identificar as condições ambientais que favorecem a ocorrência de diferentes cenários de balanço de carbono e produtividade primária.
The Amazon forest plays an important role in the terrestrial carbon balance, acting as a carbon sink through photosynthetic activity, and, at the same time, as a source of carbon via emissions from fires, wetlands and terrestrial metabolic processes. CO2 exchanges between the forest and the atmosphere can be estimated from direct observations on the surface, using the eddy covariance method. However, these observations have a small spatial representation, which cannot be extrapolated to the entire Amazon due to the heterogeneity of the carbon balance in the forest. The use of Data Science strategies can be an alternative to expand the spatial scale of carbon balance estimates, as long as the relationships between CO2 fluxes and environmental variables are known, which are often non-linear. This work aims to build machine learning models to predict the following CO2 balance metrics: net CO2 exchange (NEE between forest and atmosphere), gross primary productivity (GPP) and respiration (Re). For this, daily data of eddy covariance flux and environmental variables monitored between 2002 and 2005 in the Tapajós National Forest (FLONA-Tapajós) in the Amazon were used. As predictors, surface meteorological variables, sensible and latent heat fluxes, optical depth of aerosols and leaf area index were considered. Random Forest (RF) and Artificial Neural Network (ANN) regression models were developed. Models were also built for the classification of carbon source, sink and neutral condition scenarios. The regression models had determination coefficients (R2) between 0.33 and 0.65 for the RF models, and between 0.44 and 0.58 for the ANN models. The NEE ANN regression model correctly reproduced seasonal behavior and extreme values. Although the GPP variable reached the highest R2 values, both RF and ANN models failed to predict the extreme values of this variable. The accuracy of the classification models varied between 61% and 70%, with the RF method presenting the best performance. Between the predictive variables, those that were more relevant in the built models include: incident radiation at the top of the atmosphere, heat fluxes, leaf area index and temperature. The results obtained suggest the viability of predicting carbon fluxes in the Amazon from environmental variables, constituting the first step towards the extrapolation of local flux observations to a regional scale. The classification models allowed identifying the environmental conditions that favor the occurrence of different carbon balance and primary productivity scenarios.
The Amazon forest plays an important role in the terrestrial carbon balance, acting as a carbon sink through photosynthetic activity, and, at the same time, as a source of carbon via emissions from fires, wetlands and terrestrial metabolic processes. CO2 exchanges between the forest and the atmosphere can be estimated from direct observations on the surface, using the eddy covariance method. However, these observations have a small spatial representation, which cannot be extrapolated to the entire Amazon due to the heterogeneity of the carbon balance in the forest. The use of Data Science strategies can be an alternative to expand the spatial scale of carbon balance estimates, as long as the relationships between CO2 fluxes and environmental variables are known, which are often non-linear. This work aims to build machine learning models to predict the following CO2 balance metrics: net CO2 exchange (NEE between forest and atmosphere), gross primary productivity (GPP) and respiration (Re). For this, daily data of eddy covariance flux and environmental variables monitored between 2002 and 2005 in the Tapajós National Forest (FLONA-Tapajós) in the Amazon were used. As predictors, surface meteorological variables, sensible and latent heat fluxes, optical depth of aerosols and leaf area index were considered. Random Forest (RF) and Artificial Neural Network (ANN) regression models were developed. Models were also built for the classification of carbon source, sink and neutral condition scenarios. The regression models had determination coefficients (R2) between 0.33 and 0.65 for the RF models, and between 0.44 and 0.58 for the ANN models. The NEE ANN regression model correctly reproduced seasonal behavior and extreme values. Although the GPP variable reached the highest R2 values, both RF and ANN models failed to predict the extreme values of this variable. The accuracy of the classification models varied between 61% and 70%, with the RF method presenting the best performance. Between the predictive variables, those that were more relevant in the built models include: incident radiation at the top of the atmosphere, heat fluxes, leaf area index and temperature. The results obtained suggest the viability of predicting carbon fluxes in the Amazon from environmental variables, constituting the first step towards the extrapolation of local flux observations to a regional scale. The classification models allowed identifying the environmental conditions that favor the occurrence of different carbon balance and primary productivity scenarios.
Descrição
Citação
BAUER, L. O. Predição de trocas de carbono entre a biosfera e a atmosfera na FLONA-Tapajós a partir de variáveis ambientais. 2022. 95 f. Dissertação (Mestrado em Análise Ambiental Integrada) - Instituto de Ciências Ambientais, Químicas e Farmacêuticas, Universidade Federal de São Paulo, Diadema, 2022.
Palavras-chave
Análise Ambiental Integrada, Amazônia, Ciência dos dados, Aprendizagem de máquina, Redes Neurais Artificiais, Interação biosfera-atmosfera, Fluxos de CO2, Integrated Environmental Analysis, Amazon rainforest, Data science, Machine learning, Artificial Neural Networks, Biosphere-atmosphere interactions, CO2 fluxes