@PhDThesis{Amaral:2019:DePaMi,
author = "Amaral, Lia Martins Costa do",
title = "Development of a passive microwave-based satellite precipitation
estimation algorithm for Brazil",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2019",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2019-05-20",
keywords = "Satellite, precipitation, artificial neural networks, passive
microwave, GPM Microwave Imager (GMI), observational database.
precipita{\c{c}}{\~a}o por sat{\'e}lite, redes neurais
artificais, microondas passivo, GMI, DPR, CMB, banco de dados
observacional.",
abstract = "In order to develop a passive microwave-based satellite
precipitation estimation algorithm optimized for Brazil, this work
was divided in two parts. The first part consisted in extending
the cloud-radiation database used as a priori information for the
Cloud Dynamics and Radiation Database (CDRD) Bayesian algorithm in
order to include the cloud resolving model simulations
representative of brazilian rainfall regimes. Simulations of
microphysical, dynamical and meteorological profiles were then
generated using the University of Wisconsin - Nonhydrostatic
Modeling System and the brightness temperature (TB) simulations
were generated using the Radiative Transfer Equation Modeling
System for the CHUVA (Amazon and Vale) golden cases and compared
with observed TB. The results demonstrated that the simulations
detected perturbations in the TB fields (in space and time)
however in terms of the range of temperature values, the model did
not reproduce the lowest values of TB that were present in the
observations. The model also seemed to struggle with the riming
process on graupel formation, providing small amounts of graupel
content. These results demonstrated that the models needed
adjustments to be able to describe the regional features of TB
across a wide range of meteorological systems in Brazil. For these
reasons, the second part of the work was developed by making use
of an observational database from the sensors GPM Microwave Imager
and Dualfrequency Precipitation Radar (GMI/DPR-CMB) in order to
develop a screening of precipitation and rainfall retrieval
algorithm over Brazil, based on artificial neural networks (ANN)
and called Neural Network IMplementation of the Brazilian
MUltilayer Perceptron for Screening and precipitation retrieval
(NNIMBUS). The precipitation screening proved to be very effective
in both detecting larger systems and smaller or isolated systems.
Regarding the GMI/DPR-CMB validation dataset, the screening
performed well, with an accuracy of 0.95, POD of 0.80, FAR of 0.39
and bias of 1.34. When compared to the Goddard profiling algorithm
(GPROF) the screening still had good performance, however with
slightly smaller scores. It was observed that through the
comparison maps with GPROF the NNIMBUS can detect agglomerates
very similarly, however it does not detect the borders of the
systems very well. This behavior might be associated with the
precipitation thresholds that were configured with the training
dataset (0.2 a 60 mm/h), which might be leading more stratiform
regions of the systems to go undetected. The rainfall retrieval
model also performed well when compared to the GMI/DPR-CMB
observations, with an MAE of 4.19, standard deviation of 3.23 and
RMSE of 5.59 for the validation dataset. Analyzing the rain rate
classes, the retrieval tends to underestimate classes between 0.2
and 1 mm/h, overestimate classes between 1 and 10 mm/h and
underestimate classes greater than 10 mm/h. These features can be
associated with the input dataset distribution, as well as with
the criteria applied in data cleaning process. RESUMO: Com
objetivo de desenvolver um algoritmo de estimativa de
precipita{\c{c}}{\~a}o por sat{\'e}lite baseado em microondas
passivo otimizado para o Brasil, este trabalho foi dividido em
duas partes. A primeira parte consistiu em estender o banco de
dados de radia{\c{c}}{\~a}o de nuvens usado como
informa{\c{c}}{\~a}o a priori para o algoritmo Bayesiano Cloud
Dynamics and Radiation Database (CDRD), a fim de incluir as
simula{\c{c}}{\~o}es representativas dos regimes de
precipita{\c{c}}{\~a}o do Brazil. Simula{\c{c}}{\~o}es dos
perfis microf{\'{\i}}sicos, din{\^a}micos e meteorol{\'o}gicos
foram geradas usando o University of Wisconsin Nonhydrostatic
Modeling System e as simula{\c{c}}{\~o}es de temperatura de
brilho (TB) foram geradas usando o Radiative Transfer Equation
Modeling System para os sitemas precipitantes observados durante o
projeto CHUVA (campanhas do Vale do Para{\'{\i}}ba e Manaus). Os
resultados demonstraram que as simula{\c{c}}{\~o}es detectaram
as perturba{\c{c}}{\~o}es nos campos de TB (no espa{\c{c}}o e
no tempo), por{\'e}m em termos do intervalos de TB, o modelo
n{\~a}o reproduziu os menores valores de TB presentes nas
observa{\c{c}}{\~o}es. O modelo aparentou ter dificuldade em
gerar o processo de forma{\c{c}}{\~a}o de graupel, gerando
pequenas valores de conte{\'u}do de graupel. Esses resultados
demonstraram que os modelos precisavam de ajustes para poder
descrever as caracter{\'{\i}}sticas regionais da TB para ampla
gama de sistemas meteorol{\'o}gicos no Brasil. Por estas
raz{\~o}es, a segunda parte do trabalho consistiu no
desenvolvimento de um algoritmo de redes neurais artificais
(denominado Neural Network Implementation of the Brazilian
Multilayer Perceptron for Screening and precipitation retrieval
(NNIMBUS), tanto para detec{\c{c}}{\~a}o da {\'a}rea
precipitante screening como para recupera{\c{c}}{\~a}o da
intensidade da precipita{\c{c}}{\~a}o, utilizando um banco de
dados observacionais provindos dos sensores GPM Microwave Imager e
Dual-frequency Precipitation Radar (GMI/DPR-CMB). A
detec{\c{c}}{\~a}o de precipita{\c{c}}{\~a}o (screening)
provou ser muito eficaz na detec{\c{c}}{\~a}o de sistemas
maiores e sistemas menores ou isolados. Em rela{\c{c}}{\~a}o ao
conjunto de dados de valida{\c{c}}{\~a}o do GMI/DPR-CMB, o
algoritmo apresentou bom desempenho, com acur{\'a}cia de 0,95,
POD de 0,80, FAR de 0,39 e vi{\'e}s de 1,34. Quando comparado ao
algoritmo Goddard profiling algorithm (GPROF), a
detec{\c{c}}{\~a}o de precipita{\c{c}}{\~a}o ainda apresentava
bom desempenho, por{\'e}m com estat{\'{\i}}sticas ligeiramente
menores. Atrav{\'e}s dos mapas de compara{\c{c}}{\~a}o com o
GPROF, foi poss{\'{\i}}vel perceber que o NNIMBUS consegue
detectar os aglomerados de forma muito semelhante, por{\'e}m
n{\~a}o detecta muito bem as bordas dos sistemas. Esse
comportamento pode estar associado aos limiares de
precipita{\c{c}}{\~a}o que foram configurados com o conjunto de
dados de treinamento (0,2 a 60 mm/h), o que pode estar levando a
que regi{\~o}es mais estratiformes dos sistemas n{\~a}o sejam
detectadas. O modelo de recupera{\c{c}}{\~a}o da
precipita{\c{c}}{\~a}o tamb{\'e}m teve um bom desempenho quando
comparado com as observa{\c{c}}{\~o}es GMI/DPR-CMB, com um MAE
de 4,19, desvio padr{\~a}o de 3,23 e RMSE de 5,59 para o conjunto
de dados de valida{\c{c}}{\~a}o. Analisando as classes de taxa
de chuva, a recupera{\c{c}}{\~a}o tende a subestimar as classes
entre 0,2 e 1 mm/h, superestimar as classes entre 1 e 10 mm/h e
subestimar as classes acima de 10 mm/h. Essas
caracter{\'{\i}}sticas podem estar associadass {\`a}
distribui{\c{c}}{\~a}o do conjunto de dados de entrada, bem como
aos crit{\'e}rios aplicados no processo de limpeza de dados.",
committee = "Coelho, Simone Marilene Sievert da Costa (presidente) and Vila,
Daniel Alejandro (orientador) and Machado, Luiz Augusto Toledo and
Panegrossi, Giulia and Mattos, Enrique Vieira",
englishtitle = "Desenvolvimento de um algoritmo de estimativa de
precipita{\c{c}}{\~a}o baseada em microondas passivo para o
Brasil",
language = "en",
pages = "139",
ibi = "8JMKD3MGP3W34R/3TADRES",
url = "http://urlib.net/ibi/8JMKD3MGP3W34R/3TADRES",
targetfile = "publicacao.pdf",
urlaccessdate = "19 abr. 2024"
}