@PhDThesis{Gleriani:2004:ReNeAr,
author = "Gleriani, Jos{\'e} Marinaldo",
title = "Redes neurais artificiais para classifica{\c{c}}{\~a}o
espectro-temporal de culturas agr{\'{\i}}colas",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2004",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2004-08-30",
keywords = "identifica{\c{c}}{\~a}o de culturas agr{\'{\i}}culas,
an{\'a}lise multitemporal, {\'{\i}}ndice de
vegeta{\c{c}}{\~a}o da diferen{\c{c}}a normalizada, redes
neurais, fenologia, crop identification, temporal resolution,
normalized difference vegetation index, neural nets, phenology.",
abstract = "Investigou-se nesse trabalho uma nova metodologia de
classifica{\c{c}}{\~a}o de cenas agr{\'{\i}}colas em imagens
digitais. As cenas agr{\'{\i}}colas possuem uma complexidade
intr{\'{\i}}nseca causada pela desuniformidade fenol{\'o}gica
encontrada em uma cena, al{\'e}m da perda de particularidades
espectrais quando imageadas pelos sensores orbitais de bandas
largas. Visando solucionar esse problema, foi analisada uma
metodologia onde um pixel {\'e} analisado de maneira
cont{\'{\i}}nua no tempo, e a espectro-temporalidade obtida
{\'e} analisada atrav{\'e}s de redes neurais. Dez imagens do
sensor ETM+ (Enhanced Thematic Mapper Plus) da {\'o}rbita/ponto
220/74, do ano de 2002 da regi{\~a}o de Miguel{\'o}polis (SP)
foram utilizadas. Estas imagens foram retificadas
radiometricamente para a uniformiza{\c{c}}{\~a}o dos efeitos
atmosf{\'e}ricos e classificadas atrav{\'e}s de perceptrons de
m{\'u}ltiplas camadas treinados com o algoritmo de
retropropaga{\c{c}}{\~a}o do erro (RPE); outra rede utilizada
foi a rede de Fun{\c{c}}{\~o}es de Base Radial (FBR), al{\'e}m
do classificador Gaussiano de m{\'a}xima verossimilhan{\c{c}}a.
Foram utilizados como par{\^a}metros de entrada as bandas 3, 4 e
5, e o {\'{\i}}ndice NDVI como indicador de varia{\c{c}}{\~a}o
de IAF ({\'{\I}}ndice de {\'A}rea Foliar). As
classifica{\c{c}}{\~o}es foram supervisionadas tendo 6 classes
agr{\'{\i}}colas: feij{\~a}o1, feij{\~a}o2, milho, sorgo, cana
colhida no ano e cana de ano e meio. Foram testados diferentes
par{\^a}metros estat{\'{\i}}sticos para alimentar as redes como
a m{\'e}dia e/ou desvio padr{\~a}o de janelas com 3x3 pixels, em
tr{\^e}s combina{\c{c}}{\~o}es diferentes: m{\'e}dia e desvio
padr{\~a}o das bandas 3, 4 e 5 e do NDVI; m{\'e}dia do NDVI e
m{\'e}dia e desvio padr{\~a}o das bandas 3, 4 e 5 e, por
{\'u}ltimo, somente os arquivos de m{\'e}dia das bandas e do
NDVI. A melhor combina{\c{c}}{\~a}o de par{\^a}metros foi a
utiliza{\c{c}}{\~a}o apenas dos arquivos de m{\'e}dia, uma vez
que o uso do desvio padr{\~a}o introduziu ru{\'{\i}}do na
classifica{\c{c}}{\~a}o. Ap{\'o}s a escolha da melhor
combina{\c{c}}{\~a}o de par{\^a}metros estat{\'{\i}}sticos,
analisou-se, atrav{\'e}s da classifica{\c{c}}{\~a}o temporal, o
desempenho dos algoritmos RPE, FBR e o MaxVer. Numa an{\'a}lise
posterior, executaram-se com esses tr{\^e}s algoritmos,
classifica{\c{c}}{\~o}es de {\'u}nica data, que foram
confrontadas com a classifica{\c{c}}{\~a}o temporal. Por
{\'u}ltimo, testou-se a toler{\^a}ncia das redes neurais a dados
falhos, simulando-se a perda alternada de imagens. Essas imagens
foram suprimidas e substitu{\'{\i}}das pela m{\'e}dia entre a
imagem anterior e a posterior {\`a} data considerada. Os
desempenhos das classifica{\c{c}}{\~o}es foram analisados
atrav{\'e}s de procedimentos de estat{\'{\i}}stica kappa e
kappa condicional, este {\'u}ltimo permitiu verificar o
desempenho dos classificadores e a influ{\^e}ncia da
temporalidade para cada classe espec{\'{\i}}fica. Na
an{\'a}lise dos classificadores, o algoritmo de RPE apresentou um
valor de kappa superior {\`a} rede FBR e ao MaxVer; por{\'e}m
sem diferen{\c{c}}a significativa. A simula{\c{c}}{\~a}o de
dados falhos, resultou numa queda n{\~a}o significativa do kappa,
mas a classe sorgo desapareceu do mapa tem{\'a}tico final. O
kappa condicional mostrou que a temporalidade na
caracteriza{\c{c}}{\~a}o das culturas agr{\'{\i}}colas {\'e}
relevante principalmente com a rede RPE, embora a melhora na
classifica{\c{c}}{\~a}o nem sempre ocorra simultaneamente em
rela{\c{c}}{\~a}o aos erros de omiss{\~a}o e comiss{\~a}o de
cada classe. A {\'u}nica classe que n{\~a}o se beneficiou com o
uso da temporalidade foi a classe cana de ano. Essa classe {\'e}
de dif{\'{\i}}cil defini{\c{c}}{\~a}o do vetor
espectro-temporal pela a{\c{c}}{\~a}o antr{\'o}pica que pode
ocorrer em sete meses ao longo do ano. No entanto, a
classifica{\c{c}}{\~a}o com uma {\'u}nica data, em meados de
abril, mostrou ser bastante satisfat{\'o}ria. A an{\'a}lise
espectro-temporal de cenas agr{\'{\i}}colas processada
atrav{\'e}s de redes neurais {\'e} promissora em
rela{\c{c}}{\~a}o aos tradicionais m{\'e}todos de
classifica{\c{c}}{\~a}o. ABSTRACT: This work aimed at
investigating a new classification methodology for agricultural
scenes in digital images. Agricultural scenes are intrinsically
complex due to phenological differences found in the scene and to
the loss of spectral particularities when surveyed by broad-band
orbital sensors. In order to solve this problem, a new methodology
is presented, where the pixel is seen as a continuum in time and
the spectral-temporality is analyzed using neural networks. Ten
ETM+ images, path/row 220/74 of Miguel{\'o}polis-SP, Brazil, from
the winter of 2002 were used. These images were radiometricaly
corrected to uniform the atmospheric effects and classified by a
multilayer perceptron trained with the backpropagation error (BPE)
algorithm; another neural network used was radial basis function
(FBR), besides the Maximum Likelihood Gaussian classifier
(MaxVer). The input parameters were bands 3, 4 and 5 and the NDVI
(Normalized Difference Vegetation Index) as an LAI (Leaf Area
Index) variation indicator. Supervised classifications were used
with six agricultural classes: beans1, beans2, corn, sorghum,
one-year sugarcane and one-year-and-half sugarcane. Different ways
of feeding the network with the average and/or standard deviation
of 3x3 pixel windows were tried with three different combinations:
average and standard deviation of bands 3, 4 and 5 and of NDVI;
average of NDVI and average and standard deviation of bands 3, 4
and 5; and only the files of average of the bands and the NDVI.
The best combinations parameters was the use only the average
files, because the standard deviation introduced noise in
classification. After choosing the best statistical parameters to
be used, the performance of the BPE, FBR and the MaxVer algorithms
were analyzed through a temporal classification. Then, the
classification within each date was carried out with these three
algorithms and the results were analyzed and compared against the
temporal classification of each algorithm. At last, the tolerance
of the neural network was tested for missing data, simulating the
loss of images from every other date. These images were suppressed
and substituted by the average between the preceding and the
posterior images to the considered date. The performance of these
classifications was tested using kappa and conditional kappa
statistics; this last test allowed the evaluation of the
performance of the classifiers and of the temporal trend of each
specific class. Results for the statistical parameters showed that
using only the files of average is enough to represent the
classes, as the standard deviation introduces noise to the
classification. The BPE algorithm presented a higher kappa value
than FBR network and MaxVer algorithms, but withou significative
difference; however without statistical significance. The
simulation of missing data caused no significant decrease on kappa
statistics, but the class sorghum was suppressed from the final
thematic map. The conditional kappa showed that the use of
temporal characteristics of the data in the classification of
agricultural crops is relevant, mainly with the BPE network,
although the improvement in the classification is not always
simultaneous in relation to the commission and omissions errors of
each class. The only class that did not show an improvement with
the temporal characteristic was the one-year sugarcane. In this
class the spectral-temporal vector is difficult to define due to
tillage practices that may occur any time during seven months
throughout the year. Meanwhile, the classification using only one
date from April showed rather satisfactory. The spectral-temporal
analysis of the agricultural scenes by neural network is promising
in comparison with traditional classification methods.",
committee = "Fonseca, Leila Maria Garcia (presidente) and Epiphanio, Jos{\'e}
Carlos Neves (orientador) and Silva, Jos{\'e} Dem{\'{\i}}sio
Sim{\~o}es da (orientador) and Valeriano, M{\'a}rcio de Morisson
and Vettorazzi, Carlos Alberto and Antunes, Mauro Antonio Homem",
copyholder = "SID/SCD",
englishtitle = "Artificial neural networks to spectral-temporal classification of
agricultural crops",
language = "pt",
pages = "212",
ibi = "6qtX3pFwXQZ3P8SECKy/DFLzh",
url = "http://urlib.net/ibi/6qtX3pFwXQZ3P8SECKy/DFLzh",
targetfile = "paginadeacesso.html",
urlaccessdate = "06 maio 2024"
}