@MastersThesis{Miranda:2023:PiClLa,
author = "Miranda, Mateus de Souza",
title = "AI4LUC: pixel-based classification of land use and land cover via
deep learning and a Cerrado image dataset",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2023",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2023-03-28",
keywords = "pixel-based image classification, deep learning, Cerrado,
CBERS-4A, classifica{\c{c}}{\~a}o de imagem baseada em pixel,
aprendizado profundo, Cerrado, CBERS-4A.",
abstract = "The Cerrado biome is known for the biodiversity of flora, as well
as for its potential in agricultural production. Its landscapes of
land use and land cover (LULC) are monitored in order to analyze
and understand the social, economic, and environmental aspects
related to causative factors and impacts of these activities.
There are many efforts by the Remote Sensing (RS) community for
employing machine learning (ML) or deep learning (DL) techniques
aiming to improve classification tasks, in terms of either
pixel-based classification or contextual classification. However,
a few datasets containing images with high spatial resolution,
representativeness, and a huge number of samples about the Cerrado
biome are available. For supervised learning of either DL or ML
models, dataset samples must be labeled. This procedure currently
relies on manual execution, demanding significant time and
attention. For instance, it involves generating and labeling
reference masks, where specific pixels indicate the class to which
they belong in the segment. Driven by these motivations, this
masters dissertation strives to make a valuable contribution to
the field of pixel-based classification, specifically focusing on
semantic segmentation of Land Use and Land Cover (LULC) using deep
learning techniques applied to a dataset of satellite images from
the Cerrado region. To achieve this objective, a novel approach
named Artificial Intelligence for Land Use and Land Cover
Classification (AI4LUC) is introduced. Thus, a dataset regarding
the Cerrado biome was created, called CerraData, amounting to
unlabeled 2.5 million patches with a height and width of 256
pixels, and two meters of spatial resolution. The spectral bands
were obtained from the Wide Panchromatic and Multispectral Camera
(WPM) of the China-Brazil Earth Resources-4A (CBERS-4A) satellite.
From this dataset, two novel labeled versions were designed.
Furthermore, a novel convolutional neural network (CNN) called
CerraNetv3 has been developed to enhance the pixel-based
classification task. CerraNetv3, along with Google DeepLabv3plus,
collaboratively contributes to this endeavor. Additionally, an
innovative technique has been introduced to automate the
generation and labeling of reference masks. By leveraging the
capabilities of CerraNetv3, these reference masks are utilized to
facilitate the training process of DeepLabv3plus for pixel-based
classification. AI4LUC was subjected to a comparative analysis
with other related approaches in the domain of semantic
segmentation and contextual classification to assess its
viability. The findings revealed that CerraNetv3 achieved the
highest performance in the contextual classification experiment,
attaining an impressive F1-score of 0.9289. As for the automatic
mask generation and labeling method, it yielded an overall score
of 0.6738, with F1-score metrics. In contrast, DeepLabv3plus
obtained significantly lower scores of 0.2805 for the same metric.
The lower scores of the mask generation method can be attributed
to occasional deficiencies in the quality of generated masks,
resulting in mislabeling by the CerraNetv3 classifier.
Consequently, DeepLabv3plus also exhibited suboptimal performance.
RESUMO: O bioma Cerrado {\'e} conhecido pela biodiversidade da
flora, bem como pelo seu potencial na produ{\c{c}}{\~a}o
agr{\'{\i}}cola. Suas paisagens de uso e cobertura da terra
(LULC) s{\~a}o monitoradas a fim de analisar e compreender os
aspectos sociais, econ{\^o}micos e ambientais relacionados aos
fatores causadores e impactos dessas atividades. Existem muitos
esfor{\c{c}}os da comunidade de Sensoriamento Remoto (SR) para
empregar t{\'e}cnicas de aprendizado de m{\'a}quina (AM) ou
aprendizado profundo (AP) com o objetivo de melhorar as tarefas de
classifica{\c{c}}{\~a}o, seja em termos de
classifica{\c{c}}{\~a}o baseada em pixels ou
classifica{\c{c}}{\~a}o contextual. No entanto, poucos conjuntos
de dados contendo imagens com alta resolu{\c{c}}{\~a}o espacial,
representatividade e um grande n{\'u}mero de amostras sobre o
bioma Cerrado est{\~a}o dispon{\'{\i}}veis. Para aprendizado
supervisionado de modelos AP ou AM, as amostras de conjunto de
dados devem ser rotuladas. Este procedimento atualmente depende de
execu{\c{c}}{\~a}o manual, exigindo muito tempo e
aten{\c{c}}{\~a}o. Por exemplo, a gera{\c{c}}{\~a}o e
rotulagem de m{\'a}scaras de refer{\^e}ncia, onde cada pixel
indicam a classe a que pertencem no segmento. Impulsionada por
essas motiva{\c{c}}{\~o}es, esta disserta{\c{c}}{\~a}o de
mestrado visa contribuir para o campo da classifica{\c{c}}{\~a}o
baseada em pixels, focando especificamente na
segmenta{\c{c}}{\~a}o sem{\^a}ntica do uso e cobertura da Terra
(LULC) usando t{\'e}cnicas de AP aplicadas a um conjunto de dados
de imagens de sat{\'e}lite do Cerrado. Para alcan{\c{c}}ar este
objetivo, uma nova metodologia, denominada Artificial Intelligence
for Land Use and Land Cover Classification (AI4LUC), {\'e}
apresentada. Assim, foi criado um conjunto de dados referente ao
bioma Cerrado, denominado CerraData, totalizando 2,5 milh{\~o}es
de manchas n{\~a}o rotuladas com altura e largura de 256 pixels e
dois metros de resolu{\c{c}}{\~a}o espacial. As bandas
espectrais foram obtidas da Wide Panchromatic and Multispectral
Camera (WPM) do sat{\'e}lite CBERS-4A. A partir deste conjunto de
dados, duas novas vers{\~o}es rotuladas foram projetadas.
Al{\'e}m disso, uma nova rede neural convolucional (CNN) chamada
CerraNetv3 foi desenvolvida para tarefa de
classifica{\c{c}}{\~a}o contextual. Esta rede foi introduzida a
no m{\'e}todo para automatizar a gera{\c{c}}{\~a}o e rotulagem
de m{\'a}scaras de refer{\^e}ncia, as quais s{\~a}o utilizadas
para o treinamento do DeepLabv3plus. AI4LUC foi submetido a uma
an{\'a}lise comparativa com outras abordagens no dom{\'{\i}}nio
da segmenta{\c{c}}{\~a}o sem{\^a}ntica e
classifica{\c{c}}{\~a}o contextual para avaliar a sua
viabilidade. Os resultados revelaram que o CerraNetv3
alcan{\c{c}}ou o melhor desempenho no experimento de
classifica{\c{c}}{\~a}o contextual, atingindo de 0,9289 com
F1-score. Quanto {\`a} gera{\c{c}}{\~a}o autom{\'a}tica de
m{\'a}scara e ao m{\'e}todo de rotulagem, obteve uma
pontua{\c{c}}{\~a}o geral de 0,6738, com F1-score. As
pontua{\c{c}}{\~o}es mais baixas desse m{\'e}todo podem ser
associadas a qualidade das m{\'a}scaras geradas, resultando em
rotulagem incorreta pelo classificador CerraNetv3.
Consequentemente, o DeepLabv3plus obteve 0,2805, desempenho abaixo
do ideal esperado.",
committee = "Shiguemori, {\'E}lcio Hideiti (presidente) and Santiago
J{\'u}nior, Valdivino Alexandre de (orientador) and K{\"o}rting,
Thales Sehn (orientador) and Escada, Maria Isabel Sobral and Papa,
Jo{\~a}o Paulo",
englishtitle = "AI4LUC: classifica{\c{c}}{\~a}o baseada em pixels do uso e
cobertura da terra considerando um conjunto de imagens do
Cerrado",
language = "en",
pages = "77",
ibi = "8JMKD3MGP3W34T/48QQB65",
url = "http://urlib.net/ibi/8JMKD3MGP3W34T/48QQB65",
targetfile = "publicacao.pdf",
urlaccessdate = "04 maio 2024"
}