TY - GEN
T1 - Detecting Atlantic forest patches applying GEOBIA and data mining techniques
AU - Girolamo Neto, C.D.
AU - Pessoa, A.C.M.
AU - Körting, T.S.
AU - Fonseca, L.M.G.
N1 - Conference code: 6
PY - 2016/9/14
Y1 - 2016/9/14
N2 - Brazilian Atlantic Forest is one of the most devastated tropical forests in the world. Considering that approximately only 12% of its original extent still exists, studies in this area are highly relevant. In this context, this study maps the land cover of Atlantic Forest within the Protected Area of ‘Macaé de Cima’, in Rio de Janeiro State, Brazil, combining GEOBIA and data mining techniques on an OLI/Landsat-8 image. The methodology proposed in this work includes the following steps: (a) image pan-sharpening; (b) imagesegmentation; (c) feature selection; (d) classification and (e) model evaluation. A total of 15 features, including spectral information, vegetation indices and principal components were used to distinguish five patterns, including Water, Natural forest, Urban area, Bare soil/pasture and Rocky mountains. Features were selected considering well-known algorithms, such as Wrapper, the Correlation Feature Selection and GainRatio. Following, Artificial Neural Networks, Decision Trees and Random Forests classification algorithms were applied to the dataset. The best results were achieved by Artificial Neural Networks, when features were selected through the Wrapper algorithm. The global classification accuracy obtained was of 98.3%. All the algorithms presented great recall and precision values for the Natural forest, however the patterns of Urban area and Bare soil/pastures presented higher confusion.
AB - Brazilian Atlantic Forest is one of the most devastated tropical forests in the world. Considering that approximately only 12% of its original extent still exists, studies in this area are highly relevant. In this context, this study maps the land cover of Atlantic Forest within the Protected Area of ‘Macaé de Cima’, in Rio de Janeiro State, Brazil, combining GEOBIA and data mining techniques on an OLI/Landsat-8 image. The methodology proposed in this work includes the following steps: (a) image pan-sharpening; (b) imagesegmentation; (c) feature selection; (d) classification and (e) model evaluation. A total of 15 features, including spectral information, vegetation indices and principal components were used to distinguish five patterns, including Water, Natural forest, Urban area, Bare soil/pasture and Rocky mountains. Features were selected considering well-known algorithms, such as Wrapper, the Correlation Feature Selection and GainRatio. Following, Artificial Neural Networks, Decision Trees and Random Forests classification algorithms were applied to the dataset. The best results were achieved by Artificial Neural Networks, when features were selected through the Wrapper algorithm. The global classification accuracy obtained was of 98.3%. All the algorithms presented great recall and precision values for the Natural forest, however the patterns of Urban area and Bare soil/pastures presented higher confusion.
U2 - 10.3990/2.379
DO - 10.3990/2.379
M3 - Conference contribution
SN - 978-90-365-4201-2
BT - Proceedings of GEOBIA 2016 : Solutions and synergies, 14-16 September 2016, Enschede, Netherlands
A2 - Kerle, N.
A2 - Gerke, M.
A2 - Lefevre, S.
PB - University of Twente, Faculty of Geo-Information Science and Earth Observation (ITC)
CY - Enschede
T2 - 6th International Conference on Geographic Object-Based Image Analysis, GEOBIA 2016
Y2 - 14 September 2016 through 16 September 2016
ER -