@PhDThesis{Neves:2021:HiMaBr,
author = "Neves, Alana Kasahara",
title = "Hierarchical mapping of Brazilian Savanna (Cerrado) physiognomies
based on Deep Learning",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2021",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2021-03-22",
keywords = "spatial context, high spatial resolution image, GEOBIA, semantic
segmentation, convolutional neural network, contexto espacial,
imagem de alta resolu{\c{c}}{\~a}o espacial,
segmenta{\c{c}}{\~a}o sem{\^a}ntica, rede neural
convolucional.",
abstract = "The Brazilian Savanna, also known as Cerrado, is considered one of
the global hotspots for biodiversity conservation and plays an
important role as carbon stock, due to its above and below-ground
biomass. The Cerrado vegetation is composed by a mosaic of
ecosystems, which comprises since natural grasslands until dense
forests. There is a vegetation gradient with a wide variation in
structure, density and biomass, which generates several types of
vegetation, known as physiognomies. According to the Ribeiro and
Walter classification system, there are three major groups of
ecosystems (Grassland, Savanna and Forest), which can be divided
into 11 physiognomies and 14 additional sub-types of
physiognomies, resulting in 25 physiognomic types. Monitoring the
Cerrado vegetation cover in a large scale, using Remote Sensing
imagery, is still a challenge due to the high spatial and temporal
variability of the vegetation types and their spectral similarity.
Two aspects of the Cerrado physiognomies are relevant to create a
novel classification method: its classification system hierarchy
and the relative context where each physiognomy occurs. Two
classification techniques that considers the spatial context have
been used in the Remote Sensing field: GEOBIA and Deep Learning.
Thus, the general objective of this study is to develop and
evaluate a novel method based on Deep Learning to hierarchically
classify the Cerrado physiognomies, according to the
classification system proposed by Ribeiro and Walter, in the
Bras{\'{\i}}lia National Park, a federal environmental Protected
Area. Several spectral channels were tested as input datasets to
evaluate their importance and contribution in the classification
task and all experiments used a WorldView-2 multispectral image (2
meters spatial resolution). To demonstrate the potential of Deep
Learning techniques in the Cerrado vegetation discrimination,
hierarchical and non-hierarchical GEOBIA approaches were initially
performed to classify seven physiognomies. In addition to the
spectral bands, five vegetation indices, three fractions of the
Linear Spectral Mixture Model, three components of the Tasseled
Cap transformation and six texture features were used as features.
Compared to a GEOBIA non-hierarchical approach, the GEOBIA
hierarchical approach achieved an overall accuracy of 2.5
percentage points higher (66.4% and 68.9%, respectively). In the
Deep Learning approach, an adapted U-net architecture was used to
hierarchically classify the physiognomies. The dataset composed of
RGB bands plus the 2-band Enhanced Vegetation Index (EVI2)
achieved the best performance and was used to perform the
hierarchical classification. In the first level, which identified
Forest, Savanna and Grassland, the overall accuracy was 92.8%. For
detailed Savanna and Grassland physiognomies (second level of
classification), the overall accuracies were 86.1% and 85.0%,
respectively. The Bras{\'{\i}}lia National Park final map
obtained in this study has ten physiognomies: Gallery Forest,
Woodland Savanna, Typical Savanna, Shrub Savanna, Rupestrian
Savanna, Vereda, Rupestrian Grassland, Shrub Grassland, Open
Grassland and Humid Open Grassland. The misclassified areas are
mainly related to transition regions between the physiognomies.
Deep Learning techniques were able to understand and well
represent the physiognomy patterns. To the best of our knowledge,
this work was the first one that used Deep Learning to
discriminate the Cerrado physiognomies in this level of detail.
Besides, the accuracy rates obtained here outperformed other works
that applied traditional Machine Learning algorithms and GEOBIA
for this task. RESUMO: A Savana brasileira, conhecida como
Cerrado, {\'e} considerada um hotspot global para a
conserva{\c{c}}{\~a}o da biodiversidade, e exerce um importante
papel como estoque de carbono, devido {\`a} sua biomassa acima e
abaixo do solo. A vegeta{\c{c}}{\~a}o do Cerrado {\'e} composta
por um mosaico de ecossistemas, que abrange desde campos naturais
at{\'e} densas florestas. Existe um gradiente de
vegeta{\c{c}}{\~a}o com ampla varia{\c{c}}{\~a}o em estrutura,
densidade e biomassa, que geram diferentes tipos de
vegeta{\c{c}}{\~a}o, chamados de fitofisionomias. De acordo com
o sistema de classifica{\c{c}}{\~a}o proposto por Ribeiro e
Walter, existem tr{\^e}s grupos principais de ecossistemas
(Floresta, Savana e Campo), que podem ser divididos em 11
fitofisionomias e 14 subtipos adicionais, resultando em 25 tipos
de fitofisionomias. O monitoramento da vegeta{\c{c}}{\~a}o do
Cerrado em larga escala, usando imagens de sensoriamento remoto,
ainda {\'e} um desafio devido {\`a} alta variabilidade espacial
e temporal e {\`a} similaridade espectral das fitofisionomias.
Dois aspectos da vegeta{\c{c}}{\~a}o do Cerrado s{\~a}o
relevantes para a cria{\c{c}}{\~a}o de um novo m{\'e}todo de
classifica{\c{c}}{\~a}o: a hierarquia do sistema de
classifica{\c{c}}{\~a}o e o contexto espacial em que cada
fitofisionomia ocorre. Duas t{\'e}cnicas de
classifica{\c{c}}{\~a}o que consideram o contexto espacial
t{\^e}m sido utilizadas na {\'a}rea de Sensoriamento Remoto:
GEOBIA e Deep Learning. Assim, o objetivo geral deste trabalho
{\'e} desenvolver e avaliar um novo m{\'e}todo baseado em Deep
Learning para classificar hierarquicamente as fitofisionomias do
Cerrado, de acordo com o sistema de classifica{\c{c}}{\~a}o
proposto por Ribeiro e Walter, existentes no Parque Nacional do
Bras{\'{\i}}lia, uma Unidade de Conserva{\c{c}}{\~a}o federal.
V{\'a}rias bandas e atributos espectrais foram testados como
dados de entrada para avaliar suas contribui{\c{c}}{\~o}es na
classifica{\c{c}}{\~a}o e todos os experimentos usaram uma
imagem multiespectral WorldView-2 (resolu{\c{c}}{\~a}o espacial
de 2 metros). Para demonstrar o potencial das t{\'e}cnicas de
Deep Learning para discriminar a vegeta{\c{c}}{\~a}o do Cerrado,
inicialmente uma abordagem usando GEOBIA para classificar sete
fitofisionomias foi realizada. Al{\'e}m das bandas espectrais,
cinco {\'{\i}}ndices de vegeta{\c{c}}{\~a}o, tr{\^e}s
fra{\c{c}}{\~o}es do Modelo Linear de Mistura Espectral,
tr{\^e}s componentes da transforma{\c{c}}{\~a}o Tasseled Cap e
seis atributos de textura foram usados como atributos. Em
compara{\c{c}}{\~a}o com uma abordagem n{\~a}o hier{\'a}rquica
de GEOBIA, a abordagem hier{\'a}rquica de GEOBIA obteve uma
acur{\'a}cia global 2,5 pontos percentuais maior (66,4% e 68,9%,
respectivamente). Na abordagem com Deep Learning, uma arquitetura
U-net adaptada foi usada para classificar hierarquicamente as
fitofisionomias. O conjunto de dados composto pelas bandas RGB
mais o 2-band Enhanced Vegetation Index (EVI2) obteve o melhor
desempenho e foi usado para realizar a classifica{\c{c}}{\~a}o
hier{\'a}rquica. No primeiro n{\'{\i}}vel, que identificou
Floresta, Savana e Campo, a acur{\'a}cia global foi 92,8%. Para
as fitofisionomias detalhadas de Savana e Campo (segundo
n{\'{\i}}vel de classifica{\c{c}}{\~a}o), as acur{\'a}cias
globais foram de 86,1% e 85,0 %, respectivamente. O mapa final do
Parque Nacional de Bras{\'{\i}}lia obtido neste trabalho possui
dez fitofisionomias: Mata de Galeria, Cerrado Denso, Cerrado
T{\'{\i}}pico, Cerrado Ralo, Cerrado Rupestre, Vereda, Campo
Rupestre, Campo Sujo, Campo Limpo e Campo Limpo {\'U}mido. As
{\'a}reas classificadas incorretamente est{\~a}o relacionadas
principalmente a regi{\~o}es de transi{\c{c}}{\~a}o entre as
fitofisionomias. As t{\'e}cnicas de Deep Learning foram capazes
de entender e representar bem os padr{\~o}es das fitofisionomias.
At{\'e} onde sabemos, esse foi o primeiro trabalho que usou Deep
Learning para discriminar as fitofisionomias do Cerrado nesse
n{\'{\i}}vel de detalhamento. Al{\'e}m disso, as acur{\'a}cias
aqui obtidas superaram as de outros trabalhos que aplicaram
algoritmos tradicionais de aprendizado de m{\'a}quina e GEOBIA
para essa tarefa.",
committee = "Arag{\~a}o, Luiz Eduardo Oliveira e Cruz de (presidente) and
K{\"o}rting, Thales Sehn (orientador) and Fonseca, Leila Maria
Garcia (orientadora) and Sant'Anna, Sidnei Jo{\~a}o Siqueira and
Oliveira, Cleber Gonzales de and Alencar, Ane Auxiliadora Costa",
englishtitle = "Mapeamento hier{\'a}rquico das fitofisionomias da Savana
brasileira (Cerrado) baseado em Deep Learning (aprendizagem
profunda)",
language = "en",
pages = "96",
ibi = "8JMKD3MGP3W34R/44DTSUS",
url = "http://urlib.net/ibi/8JMKD3MGP3W34R/44DTSUS",
targetfile = "publicacao.pdf",
urlaccessdate = "05 maio 2024"
}