| |
|
|
| Registros recuperados : 1 | |
| Registros recuperados : 1 | |
|
|
|
Registro Completo
|
Biblioteca(s): |
Epagri-Sede. |
|
Data corrente: |
10/01/2019 |
|
Data da última atualização: |
10/01/2019 |
|
Tipo da produção científica: |
Resumo em Anais de Congresso |
|
Autoria: |
COELHO, F. F.; GIASSON, E.; CAMPOS, A. R.; COSTA, J. J. F.; COBLINSKI, J. A.; SILVA, E. B. |
|
Título: |
ARTIFICIAL NEURAL NETWORKS AND OBJECT-BASED CLASSIFICATION FOR DIGITAL SOIL MAPPING. |
|
Ano de publicação: |
2018 |
|
Fonte/Imprenta: |
In: WORLD CONGRESS OF SOIL SCIENCE, 21., 2018, Rio de janeiro. Abstracts... Viçosa: Sociedade Brasileira de Ciência do Solo, 2018. |
|
Idioma: |
Inglês |
|
Conteúdo: |
Detailed soil maps are not available in most of the Brazilian territory. Remote sensing data and
artificial intelligence techniques can be integrated into an object-based classification to improve
spatial prediction of soil class and soil map availability. This work aimed (i) to compare the spatial
predictive of soil class of Artificial Neural Network (ANN) and Decision Trees (DT), (ii) to produce
digital soil maps by an object-based classification. The study area is located in northwest Rio
Grande do Sul State, southern Brazil. The surface area is 900 km². In this area, there are four soil
orders: Oxisols, Molisols, Inceptisols and Entisols. Normalized Difference Vegetation Index (NDVI),
Normalized Difference Water Index (NDWI) and Normalized Difference Soil Index (NDSI) were
derived from Landsat 8 OLI (Operational Land Imager) sensor imagery. Fifteen terrain attributes
were derived from SRTM (Shuttle Radar Topography Mission) using RSAGA package in RStudio
environment. This spectral indices and terrain attributes were used as discriminating variables. The
multiresolution segmentation (MRS) algorithm (eCognition Developer 9.0) was used to create image
objects. The shape and compactness criterion were 0.1, the scale parameter (SP) tested were 1, 2,
5, 10, 25 and 50. Three classification repetitions with Multilayer Perceptron (ANN), Simple Cart (DT)
and J48 (DT) algorithms were performed with 4993 random samples using cross-validation
technique (5 folds) in Weka Experiment Environment. The algorithm with the greater overall
accuracy (OA) was used to object-based classification. The object-based digital soil maps accuracies
were evaluated by the agreement with a legacy soil map. The Multilayer Perceptron algorithm had
the highest OA (61%), followed by Simple Cart (59%) and J48 (52%). The MRS provided the
reduction from 999,206 pixels to 23,616 (SP 1), 9,528 (SP 2), 3,658 (SP 5), 2,350 (SP 10), 1,769 (SP
25) and 1,660 (SP 50) image objects. The object-based approach OA with original legend were 62%
(SP 1), 60% (SP 2), 57% (SP 5), 53% (SP 10), 50% (SP 25) and 45% (SP 50). The object-based
approach OA with simplified legend were 77% (SP 1), 75% (SP 2), 72% (SP 5), 69% (SP 10), 66% (SP
25) e 62% (SP 50). The classifier based on ANN had the higher OA than the classifiers based on DT.
Larger SP provides greater object?s size and smaller its number to classify thus reducing the OA.
The object-based classification is a promising approach for digital soil mapping MenosDetailed soil maps are not available in most of the Brazilian territory. Remote sensing data and
artificial intelligence techniques can be integrated into an object-based classification to improve
spatial prediction of soil class and soil map availability. This work aimed (i) to compare the spatial
predictive of soil class of Artificial Neural Network (ANN) and Decision Trees (DT), (ii) to produce
digital soil maps by an object-based classification. The study area is located in northwest Rio
Grande do Sul State, southern Brazil. The surface area is 900 km². In this area, there are four soil
orders: Oxisols, Molisols, Inceptisols and Entisols. Normalized Difference Vegetation Index (NDVI),
Normalized Difference Water Index (NDWI) and Normalized Difference Soil Index (NDSI) were
derived from Landsat 8 OLI (Operational Land Imager) sensor imagery. Fifteen terrain attributes
were derived from SRTM (Shuttle Radar Topography Mission) using RSAGA package in RStudio
environment. This spectral indices and terrain attributes were used as discriminating variables. The
multiresolution segmentation (MRS) algorithm (eCognition Developer 9.0) was used to create image
objects. The shape and compactness criterion were 0.1, the scale parameter (SP) tested were 1, 2,
5, 10, 25 and 50. Three classification repetitions with Multilayer Perceptron (ANN), Simple Cart (DT)
and J48 (DT) algorithms were performed with 4993 random samples using cross-validation
technique (5 folds) in Weka Experiment En... Mostrar Tudo |
|
Palavras-Chave: |
data mining; multilayer perception; multiresolution segmentation. |
|
Categoria do assunto: |
X Pesquisa, Tecnologia e Engenharia |
|
|
|
|
Marc: |
LEADER 03223naa a2200217 a 4500
001 1128201
005 2019-01-10
008 2018 bl uuuu u00u1 u #d
100 1 $aCOELHO, F. F.
245 $aARTIFICIAL NEURAL NETWORKS AND OBJECT-BASED CLASSIFICATION FOR DIGITAL SOIL MAPPING.$h[electronic resource]
260 $c2018
520 $aDetailed soil maps are not available in most of the Brazilian territory. Remote sensing data and artificial intelligence techniques can be integrated into an object-based classification to improve spatial prediction of soil class and soil map availability. This work aimed (i) to compare the spatial predictive of soil class of Artificial Neural Network (ANN) and Decision Trees (DT), (ii) to produce digital soil maps by an object-based classification. The study area is located in northwest Rio Grande do Sul State, southern Brazil. The surface area is 900 km². In this area, there are four soil orders: Oxisols, Molisols, Inceptisols and Entisols. Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI) and Normalized Difference Soil Index (NDSI) were derived from Landsat 8 OLI (Operational Land Imager) sensor imagery. Fifteen terrain attributes were derived from SRTM (Shuttle Radar Topography Mission) using RSAGA package in RStudio environment. This spectral indices and terrain attributes were used as discriminating variables. The multiresolution segmentation (MRS) algorithm (eCognition Developer 9.0) was used to create image objects. The shape and compactness criterion were 0.1, the scale parameter (SP) tested were 1, 2, 5, 10, 25 and 50. Three classification repetitions with Multilayer Perceptron (ANN), Simple Cart (DT) and J48 (DT) algorithms were performed with 4993 random samples using cross-validation technique (5 folds) in Weka Experiment Environment. The algorithm with the greater overall accuracy (OA) was used to object-based classification. The object-based digital soil maps accuracies were evaluated by the agreement with a legacy soil map. The Multilayer Perceptron algorithm had the highest OA (61%), followed by Simple Cart (59%) and J48 (52%). The MRS provided the reduction from 999,206 pixels to 23,616 (SP 1), 9,528 (SP 2), 3,658 (SP 5), 2,350 (SP 10), 1,769 (SP 25) and 1,660 (SP 50) image objects. The object-based approach OA with original legend were 62% (SP 1), 60% (SP 2), 57% (SP 5), 53% (SP 10), 50% (SP 25) and 45% (SP 50). The object-based approach OA with simplified legend were 77% (SP 1), 75% (SP 2), 72% (SP 5), 69% (SP 10), 66% (SP 25) e 62% (SP 50). The classifier based on ANN had the higher OA than the classifiers based on DT. Larger SP provides greater object?s size and smaller its number to classify thus reducing the OA. The object-based classification is a promising approach for digital soil mapping
653 $adata mining
653 $amultilayer perception
653 $amultiresolution segmentation
700 1 $aGIASSON, E.
700 1 $aCAMPOS, A. R.
700 1 $aCOSTA, J. J. F.
700 1 $aCOBLINSKI, J. A.
700 1 $aSILVA, E. B.
773 $tIn: WORLD CONGRESS OF SOIL SCIENCE, 21., 2018, Rio de janeiro. Abstracts... Viçosa: Sociedade Brasileira de Ciência do Solo, 2018.
Download
Esconder MarcMostrar Marc Completo |
|
Registro original: |
Epagri-Sede (Epagri-Sede) |
|
|
Biblioteca |
ID |
Origem |
Tipo/Formato |
Classificação |
Cutter |
Registro |
Volume |
Status |
Fechar
|
| Nenhum registro encontrado para a expressão de busca informada. |
|
|
