@Article{MartinezAdTuCoAlFe:2022:CoClRe,
author = "Martinez, J. A. C. and Adarme, M. X. O. and Turnes, J. N. and
Costa, Gilson A. O. P. and Almeida, Claudio Aparecido de and
Feitosa, Raul Q.",
affiliation = "{Pontif{\'{\i}}cia Universidade Cat{\'o}lica do Rio de Janeiro
(PUC-Rio)} and {Pontif{\'{\i}}cia Universidade Cat{\'o}lica do
Rio de Janeiro (PUC-Rio)} and {University of Waterloo} and
{Universidade do Estado do Rio de Janeiro (UERJ)} and {Instituto
Nacional de Pesquisas Espaciais (INPE)} and {Pontif{\'{\i}}cia
Universidade Cat{\'o}lica do Rio de Janeiro (PUC-Rio)}",
title = "A comparison of cloud removal methods for deforestation monitoring
in Amazon rainforest",
journal = "International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences",
year = "2022",
volume = "43",
number = "B3",
pages = "665--671",
month = "June",
keywords = "Cloud Removal, Deep learning, Deforestation, Optical imagery,
SAR-optical Data fusion.",
abstract = "Deforestation in tropical rainforests is a major source of carbon
dioxide emissions, an important driver of climate change. For
decades, the Brazilian government has maintained monitoring
programs for deforestation detection in the Brazilian Legal Amazon
area based on remotely sensed optical images in a protocol that
involves considerable efforts of visual interpretation. However,
the Amazon region is covered with clouds for most of the year, and
deforestation assessment can rely only on images acquired in the
dry season when cloud-free images are more likely to capture. One
possibility to lessen that restriction and enable deforestation
detection throughout the year is to synthesize cloud-free optical
images from corresponding SAR images, which are only marginally
influenced by atmospheric conditions. This work compares a set of
such image synthesis methods, considering deforestation detection
in the Amazon forest as the target application. Specifically, we
evaluate three deep learning methods for cloud removal in
Sentinel-2 images: a conditional Generative Adversarial Network
(cGAN) based on the pix2pixi architecture; an extension of that
method, which uses atrous convolutions (Atrous cGANi) to enhance
fine image details; and a non-generative method (DSen2-CRi) based
on residual networks. In the evaluation, we assess both the
quality of the generated images and the accuracy obtained when
performing deforestation detection from those images. We further
compare those methods with an image aggregation tool available in
Google Earth Engine (GEE Tooli), which creates cloud-free mosaics
from sequences of images acquired at nearby dates. In this study,
we considered two sites in the Brazilian Amazon, characterized by
distinct vegetation and deforestation patterns. In terms of the
quality metrics and classification accuracy, the Atrous cGANi was
the best performing deep learning method. The GEE Tooli
outperformed all those methods when dealing with images from the
dry season but turned out to be the poorest performing method in
the wet season.",
doi = "10.5194/isprs-archives-XLIII-B3-2022-665-2022",
url = "http://dx.doi.org/10.5194/isprs-archives-XLIII-B3-2022-665-2022",
issn = "1682-1750",
language = "en",
targetfile = "isprs-archives-XLIII-B3-2022-665-2022.pdf",
urlaccessdate = "29 jun. 2024"
}