@InProceedings{PachecoMaSiSoShEs:2021:ImClMe,
author = "Pacheco, Fl{\'a}via Domingos and Matias, Ma{\'{\i}}ra Ramalho
and Silva, Gabriel M{\'a}ximo da and Souza, Anielli Rosane de and
Shimabukuro, Yosio Edemir and Escada, Maria Isabel Sobral",
affiliation = "{Instituto Nacional de Pesquisas Espaciais (INPE)} and {Instituto
Nacional de Pesquisas Espaciais (INPE)} and {Instituto Nacional de
Pesquisas Espaciais (INPE)} and {Instituto Nacional de Pesquisas
Espaciais (INPE)} and {Instituto Nacional de Pesquisas Espaciais
(INPE)} and {Instituto Nacional de Pesquisas Espaciais (INPE)}",
title = "Image Classification Methods Assessment for Identification of
Small-Scale Agriculture in Brazilian Amazon",
booktitle = "Proceedings...",
year = "2021",
pages = "12--19",
organization = "International Conference on Advanced Geographic Information
Systems, Applications, and Services, 13. (GEOProcessing)",
publisher = "IARIA",
keywords = "digital image processing, segmentation, land use, land cover,
smallholders, planetscope.",
abstract = "This paper aims to test different methods for image classification
focusing on small-scale agriculture in the region of Mocajuba and
Camet{\'a}, municipalities in the Northeast of Par{\'a} state,
Brazil. It is an important land use class, always ignored by
Land-Use and Land-Cover monitoring systems because of its small
size and variable spectral signature. We used an image from the
PlanetScope Surface Reflectance Mosaics (Analysis Ready) with
spatial resolution of 4.77 meters and 4 spectral bands (red,
green, blue and infra-red). After proceeding with a
multiresolution segmentation to identify image objects, two
object-oriented classification algorithms were tested: Adapted
Nearest-neighbor and C5.0 Decision trees algorithms. We selected
122 random points using the images available on Google Earth Pro
as reference to assess the accuracy of classifications.
Afterwards, confusion matrices were generated. Both methods showed
similar overall accuracy and kappa value. However, C5.0 Decision
trees reached a higher producers accuracy to small-scale
agriculture (75%) in comparison to Adapted Nearest-neighbor (65%).
The average size of the small-scale agriculture segments estimated
was less than 1 ha in both maps, showing the need to carry out
studies on scales of greater detail, preferably with images of
high spatial resolution to represent these systems properly. In
this study, C5.0 Decision trees had the best result, representing
the most suitable method for mapping small-scale agriculture in
Brazilian Amazon.",
conference-location = "Nice, France",
conference-year = "19-22 july",
isbn = "978-1-61208-871-6",
issn = "2308-393X",
language = "en",
ibi = "8JMKD3MGP3W34T/454U5F5",
url = "http://urlib.net/ibi/8JMKD3MGP3W34T/454U5F5",
targetfile = "geoprocessing_2021_1_40_30034.pdf",
urlaccessdate = "11 maio 2024"
}