@InProceedings{MarettoFonsKört:2017:DeLeTe,
author = "Maretto, Raian Vargas and Fonseca, Leila Maria Garcia and
K{\"o}rting, Thales Sehn",
affiliation = "{Instituto Nacional de Pesquisas Espaciais (INPE)} and {Instituto
Nacional de Pesquisas Espaciais (INPE)} and {Instituto Nacional de
Pesquisas Espaciais (INPE)}",
title = "Deep Learning Techniques Applied to classification of Remote
Sensing Images",
booktitle = "Anais...",
year = "2017",
organization = "Workshop dos Cursos de Computa{\c{c}}{\~a}o Aplicada do INPE,
17. (WORCAP)",
keywords = "Deep Learning, Remote Sensing, Machine Learning, Image
classification.",
abstract = "Remote Sensing (RS) techniques have become increasingly important
in data-collection tasks and location-based services. Recent
increased accessibility of new generation multispectral sensors
has improved the complexity required in the analysis techniques.
Produce efficient representations and understandings of the scenes
has become a challenging problem. To improve knowledge
representation and feature description, huge number of algorithms
have been developed considering not only the local pixel
information, but contextual information obtained from homogeneous
regions in images (K{\"O}RTING; GARCIA FONSECA; C{\^A}MARA,
2013; WA{\c{c}}LTER, 2004). However, most approaches lack on
learning efficient representations of the images, extracting only
shallow features that cannot easily represent the details of
complex real data (LECUN; BENGIO; HINTON, 2015; ZHANG; ZHANG;
KUMAR, 2016). Deep Learning (DL) techniques, which can learn
representative and discriminative features from data, has become a
hotspot in the Machine Learning community. They are composed of
multiple levels of feature extraction layers. Each level
transforms the representation of the previous level into a higher,
slightly more abstract model, mapping different levels of
abstractions and combining them to model and explore intrinsic
correlations of the data (Lecun et al., 2015). DL algorithms have
recently started to be used by the RS community, being
successfully used in several tasks, from pre-processing to
classification. Despite the great potential of these techniques,
many questions are still unknown for its use in RS applications.
The large number of bands and the way to consider the spectral
curves represent a great challenge. Only few labeled samples are
available, leading to difficulties to train the network. Images
acquired from different sensors or in different seasons have large
differences among them, leading to problems to transfer the
network knowledge between different images (ZHANG; ZHANG; KUMAR,
2016). The main goal of this work is to investigate the use of
Deep Learning based approaches for classification of remote
sensing images. We believe that designing an architecture to a
Deep Neural Network considering the particularities and
complexities of RS images, we can achieve good results for
classification. With this approach, we expect to answer some
opened questions about the use of DL in RS image analysis, filling
in some gaps in the image analysis. Therefore, the main question
we aim to answer is What is the best architecture to a Deep Neural
Network to classify high resolution remote sensing Images?. A case
study was developed in the classification of Land Cover in
Brazilian Amazon, with main focus on the deforestation. To train
the network and evaluate the results, PRODES deforestation data
was used. It is important to emphasize that although this study is
in a preliminary stage, the results are promising and reached
improvements in the accuracy of the classification.",
conference-location = "S{\~a}o Jos{\'e} dos Campos, SP",
conference-year = "20-22 nov. 2017",
language = "en",
targetfile = "Maretto_deep.pdf",
urlaccessdate = "11 maio 2024"
}