( PDF ) Remote sensing of deciduous forests: a multi temporal approach

Download PDF

ads:

REMOTE SENSING OF DECIDUOUS

FORESTS: A MULTI TEMPORAL

APPROACH

THOMAZ CHAVES DE ANDRADE OLIVEIRA

2009

ads:

Livros Grátis

http://www.livrosgratis.com.br

Milhares de livros grátis para download.

THOMAZ CHAVES DE ANDRADE OLIVEIRA

REMOTE SENSING OF DECIDUOUS FORESTS: A

MULTI TEMPORAL APPROACH

Dissertação apresentada à Universidade Federal de

Lavras, como parte das exigências do Curso de

Mestrado em Engenharia Florestal, área de

concentração em Manejo Florestal, para obtenção do

título de “Mestre”.

Orientador

Prof. Luis Marcelo Tavares de Carvalho

LAVRAS

MINAS GERAIS – BRASIL

2009

ads:

Oliveira, Thomaz Chaves de Andrade.

Remote sensing of deciduous forests: a multi-temporal approach /

Thomaz Chaves de Andrade Oliveira. – Lavras : UFLA, 2009.

41 p. : il.

Dissertação (Mestrado) – Universidade Federal de Lavras, 2009.

Orientador: Luis Marcelo Tavares de Carvalho.

Bibliografia.

1. Remote sensing. 2. Wavelets analysis. 3. MODIS. 4. HANTS.

5. NDVI. 6. Harmonic analysis. 7. Fourier. 8. Time series. I.

Universidade Federal de Lavras. II. Título.

CDD – 634.92

Ficha Catalográfica Preparada pela Divisão de Processos Técnicos da

Biblioteca Central da UFLA

SUMMARY

ABSTRACT…………………………………………………………………….5

RESUMO………………………………………………………………………..6

General Introduction……………………………..……………………….……7

CHAPTER 01: Analysis of Temporal NDVI profiles on different geographical

occurrences of Deciduous Forests in the Brazilian Cerrado……………………..9

1-Introduction…………………………………………………………..………10

2-Methods………………………………………………………………..…..…12

3-Results………………………………..………………………………………21

3-Conclusions…………………………………………………………………..23

4-References…………………………………………………………...……….24

CHAPTER 02: Mapping Deciduous forests using time series of filtered MODIS

NDVI and Neural Networks……………………………………………………26

1-|Introduction…………………………………………………………..……...27

2-Methods…………………………………………………………………..…..28

3-Results………………………………………………………………………..35

3-Conclusions…………………………………………………………………..38

4-References…………………………………………………………...……….39

General Conclusions…………………………..........................……………….42

ABSTRACT

OLIVEIRA, Thomaz Chaves de Andrade Oliveira. Mapping Deciduous

Forests: A multi-temporal approach. 2009 Dissertação (Mestrado em Manejo

Ambiental) - Universidade Federal de Lavras, Lavras, MG.¹

This work investigates whether Deciduous vegetation of Minas Gerais, in the

Cerrado Biome, have differences in geographically different areas. This was

accomplished through the use of Normalized Difference Vegetation Index

(NDVI) time series data. Its objective was to find out if these differences exist

and to quantify these differences. Four different geographical locations covered

by deciduous forests and other neighbor vegetation were chosen for the analysis.

This study was conducted in two chapters where chapter 01 quantifies the annual

shifts of phenology curves of Deciduous Forests in the different regions and

chapter 02 applies different denoising algorithms to time series of NDVI data

and compared the results of vegetation classification as input to artificial neural

networks. It was concluded that there are annual shifts in the annual curve

among deciduous forests localized in different geographical areas of Minas

Gerais state. Time signatures of NDVI can be used with success to map

Deciduous Forests, however the results do not conclude which of the two

filtering techniques best generates the vegetation signatures.

RESUMO

OLIVEIRA, Thomaz Chaves de Andrade Oliveira. Mapping Deciduous

Forests: A multi-temporal approach. 2009 Dissertação (Mestrado em Manejo

Ambiental) - Universidade Federal de Lavras, Lavras, MG.¹

Este presente trabalho tem o objetivo principal de pesquisar se as florestas

deciduais de Minas Gerais situadas em áreas geográficas distantes possuem

diferenças entre as mesmas. Isso foi possível através da utilização de séries

temporais do Índice de Vegetação de Diferença Normalizada (NDVI), para

quantificar as diferenças. Foram escolhidas quatro diferentes localizações

separadas geograficamente, onde encontram-se florestas decíduas e outros tipos

de vegetação. Este trabalho foi conduzido em dois capítulos onde o capítulo 01

quantifica o deslocamento temporal anual da curva da fenologia das florestas

deciduais de diferentes regiões. O capítulo 02 aplica diferentes algoritmos de

remoção de ruído para séries temporais de NDVI e compara os resultados

através da classificação da vegetação com a utilização das séries temporais

filtradas como entrada de dados em uma rede neural artificial. Na conclusão

geral do trabalho, pode-se concluir que: existe um deslocamento temporal na

curva anual de fenologia entre as florestas decíduas que se situam em áreas

geográficas diferentes. Assinaturas temporais de índices de vegetação NDVI

podem ser utilizadas com sucesso para mapear as florestas decíduas de

diferentes localizações, no entanto, não evidenciam a melhor entre as técnicas de

filtragem de dados.

GENERAL INTRODUCTION

The Cerrado Biome in the Minas Gerais state-Brazil is one of which is

in constant threat to deforestation. One of the phytophysiognomies within this

biome that needs special attention is the Deciduous Forest. It is characterized by

an alternating cycle of dry and wet seasons. By which 70 % of its leaves are off

in the dry season (Oliveira-Filho & Ratter, 2006). The peculiarities of this

vegetation involve special treatment when mapping land cover due to this

variation on “greenness” throughout time. Since ‘objects’ that have similar

spectral reflectance present problems when mapping with single date remote

sensing images, when mapping this vegetation, incorrect results may appear in

some locations. This work’s objective was to generate time information based

on time series of Normalized Difference Index that will help map this forest

type. The time series used in the study were collected during 2003, 2004 and

2005, all the results were conducted in four different locations.

This study was motivated by setting up the following questions surrounding the

deciduous forests:

1) Do geographically distant deciduous forests have annual shifts in their

phenological cycles during the year?

2) Can MODIS filtered NDVI (Normalized Difference Vegetation Index)

time series generate precise mapping to deciduous forests in different

regions?

3) Which of two filtering techniques (HANTS Fourier Analysis) and

Wavelet Filtering produces the best filtered time series for mapping this

phytophysiognomy.

To answer these questions this work was organized in the following

sequence:

General Introduction

CHAPTER 01 – Analysis of Temporal NDVI profiles on different

geographical occurrences of Deciduous Forests in the Brazilian Cerrado

The Objective of this work is to answer the question 01, it uses harmonic

analyses to calculate the annual shift of the phonological curves of each

geographical separated area.

CHAPTER 02 - Mapping Deciduous forests using time series of filtered

MODIS NDVI and Neural Networks

This chapter answers questions 02 and 03, where NDVI time series of

four different geographical regions were filtered with the HANTS algorithms

and Wavelet temporal algorithms, the resulting data sets were input to

classification in an artificial neural network.

General Conclusions

Chapter 01

Analysis of Temporal NDVI profiles on different geographical

occurrences of Deciduous Forests in the Brazilian Cerrado

Thomaz C. de A. Oliveira

, Luis M. T. de Carvalho

, Luciano

T. de Oliveira

, Adriana Z. Martinhago

, Fausto W. Acerbi

Júnior

1,2,3,4,5

Departamento de Ciências Florestais, Universidade

Federal de Lavras (UFLA).

Caixa Postal 3.037 – 37200-000 – Lavras – MG – Brazil

e-mail:

thomazchaves@gmail.com,

passarinho@ufla.br,

oliveiralt@yahoo.com.br,

dricazm@gmail.com,

[email protected]

(Prepared according to Elsevier Remote sensing of the Environment)

Abstract: This work investigates whether geographically distant deciduous

forests have different timing in their phenological cycle during the year. The

study was conducted in the state of Minas Gerais in the Cerrado biome. Four

different geographical locations covered by deciduous forests were chosen for

the analysis. A MODIS NDVI data set was chosen to deliver the results because

of its high time resolution. The harmonic analysis based HANTS algorithm was

used to generate the time profile of each location. The mean value of phase

from the first harmonic was used to characterize the annual shift of vegetation

phenology for each one the four regions of interest. Results show the existence

of a time gap in the annual curves of NDVI of each of the four locations.

Keywords: Remote Sensing, MODIS, NDVI, time series, HANTS, harmonic

analysis

1. Introduction

The Cerrado biome of tropical South America covers about two million

squared kilometers, representing almost 22% of the Brazilian territory. The

biome was named after the vernacular term of its predominant vegetation type, a

fairly dense woody savanna of shrubs and smalls trees. The term “cerrado”

(Portuguese for half-closed, closed, or dense) was probably applied to this

vegetation because of the difficulty of traversing it on horseback (Oliveira-Filho

& Ratter 2002). The constant threat to the Brazilian Cerrado has lead to the

necessity of strategies and measures to promote the monitoring and mapping of

this biome. The Cerrado has a large biodiversity but it’s fragmentation

throughout the years has lead to the losses of exemplars from this biome

(Oliveira 2004).

One of the phytophysiognomies within this biome that needs special

attention is the Deciduous Forest. It is characterized by an alternating cycle of

dry and wet seasons. More than 70 % of deciduous forests leaves are off in the

dry season (Oliveira-Filho, 2006). The period of dryness occurs from mid April

till September. The wet season starts in October and goes up to March. This

leads to an intensified variation of greenness in vegetation and landscape

characterization throughout time. The variation of greenness of the semi-

deciduous forests is not so intense, due to their occurrence in regions of

intensified humidity (Oliveira, 2004). These differences can be observed in

Figure 1.

Mapping land cover by means of remotely sensing data has been an area

of growing research interest throughout the past decades. Its complexity,

peculiarities and state of the art concerning computational aids and processing

routines differ a lot from past conventional cartographic tools. Developments in

computer science have aided a better information extraction from remotely

sensed images, as well as an effective use of geographical information systems

to store, analyze and present all sorts of land cover information (Carvalho,

2001).

Some objects on the Earth´s surface reflect the electro-magnetic energy

in the same way when sensed with a multi-spectral scanner. In the present case,

it is difficult to differentiate deciduous and semi-deciduous forests when leaves

are on using single date remote sensing. Nevertheless, ‘objects’ reflectance may

vary according to growth stage, phenology, humidity, atmospheric transparency,

illumination conditions etc. These characteristics led to a search for alternative

features to enable the discrimination of land cover classes with similar

reflectance behavior (Carvalho et al., 2004).

(a) (b)

Figure 1 – (a) Deciduous forest, (b) semi-deciduous forest - Source Oliveira (2004)

These features, especially temporal information, are very useful for

characterizing deciduous forests in the Cerrado biome, due to their pronounced

dynamics. This can be noticed in the official forest map of Minas Gerias state,

carried out by Carvalho (2008), (which was done with lower time resolution

Landsat TM images) that does not capture small deciduous forests fragments

present in the region of the Triângulo Mineiro, western Minas Gerais. In the

coordinates of 19º09’10’’S , 50º39’10’’W, for instance, there is a 25ha fragment

of dry Deciduous Forests on the margins of Rio Paranaíba, in fazenda Bonanza,

municipality of Santa Vitória (Oliveira-Filho et al., 1998). Research however,

suggests that remotely sensed time series data could possibly improve

misclassification and the accuracy of mapping deciduous forests (Oliveira,

2004).

According to Jensen (2000), timing is very important when attempting

to identify different vegetation types or to extract useful vegetation biophysical

information (e.g. biomass, chlorophyll, characteristics) from remote sensed data.

Multi-temporal satellite images composites are now of standard use in land

cover classification of large areas at regional and global scales (Carrão et al.

2007).

The objective of this work was to find whether geographically distant

deciduous forests have different timing in their phenological cycle during the

year, leading to an annual shift in their cycle. The answer to this question could

further improve the accuracy of forest mapping in the State of Minas Gerais.

2. Methods

2.1 Vegetation Indices

Temporal information used in this study comprised time series of

vegetation indices, viz. the Normalized Difference Vegetation Index (NDVI).

Since the 1960’s, scientists have extracted and modeled vegetation biophysical

variables using remotely sensed data. Much of the effort has gone into the

development of vegetation indices – defined as dimensionless radiometric

measures that function as indicators of relevant abundance and activity of green

vegetation, often including leaf-area-index (LAI), percentage green cover,

chlorophyll content, green biomass, and absorbed photosynthetically active

radiation (APAR). There are more than 20 vegetation indices in vigor. A

vegetation index should maximize sensitivity to plant biophysical parameters;

normalize or model external effects such as sun angle, viewing angle, and the

atmosphere for consistent spatial and temporal comparisons; normalize internal

effects such as canopy background variations. A vegetation index may

preferably couple with a measurable biophysical parameter such as biomass,

LAI, or APAR (Jensen, 2000).

Vegetation dynamics indicate important short and long-term ecological

process. Continuous temporal observations of land surface parameters using

satellite reveal seasonal and inter-annual developments. Vegetation indices have

been extensively applied to characterize the state and dynamics of vegetation, in

particular multiple NDVI datasets of the Advanced Very High Resolution

Radiometer (AVHRR) instrument used during the last 25 years (Jensen, 2000;

Coldiz et al., 2007).

Different vegetation types exhibit distinctive seasonal patterns on NDVI

variation (Yu et al., 2004). Vegetation profiles of deciduous and semi-deciduous

forests are illustrated in figure 2. In most cases, different types of vegetation

have different phenological patterns. For example, evergreen plants will have a

more steady temporal dynamics throughout the year when compared to

tropicalplants that loose their leaves (Bruce et al., 2006).

Deciduous Forests time signatures Semi-Deciduous forests time signature

Figure 2 –Denoised NDVI time series Deciduos and Semi-deciduous Forests

Spatial and temporal variability in vegetation indices arise from several

vegetation related properties, including LAI, canopy structure/architecture,

species composition, land cover type, leaf optics, canopy crown cover,

understory vegetation, and green leaf biomass (Huete et al., 2002).

2.2 The MODIS Sensor

In the present study, NDVI time series from the Moderate-resolution

Imaging Spectroradiometer (MODIS) were used. MODIS data products offer a

great opportunity for phenology-based land-cover and land use change studies

by combining characteristics of both AVHRR and Landsat, including: moderate

resolution, frequent observations, enhanced spectral resolution, and improved

atmospheric calibration (Galford et al., 2007). The AVHRR sensor was

originally designed for meteorological applications, and has only two spectral

bands (red and near-infrared) that can be used to generate spectral indices of

vegetation. The new generation MODIS sensor has a number of advantages over

AVHRR, including more spectral bands that can be used for vegetation analysis

(Yu et al., 2004).

2.3 MODIS Vegetation Indices

MODIS vegetation indices are appropriate for vegetation dynamics

studies and characterization. They are found to be sensitive to multi-temporal

(seasonal) vegetation variations and to be correlated with LAI across a range of

canopy structure, species composition, lifeforms, and land cover types. The

MODIS-NVDI demonstrates a good dynamic range and sensitivity for

monitoring and assessing spatial and temporal variations in vegetation amount

and condition. The seasonal profiles provided by the MODIS-NDVI outperform

in sensitivity and fidelity the equivalent AVHRR-NDVI profiles, particularly

when the atmosphere has a relatively high content of water vapor (Huete et al.,

2002).

2.4 Data set

Due to the widespread occurrence of deciduous forests in the state of

Minas Gerais and its large extent, 586.528 km², four different areas of interest

were chosen so that time signatures of geographically separated forests could be

compared. These locations were primarily chosen because of known

occurrences of Deciduous forests according to the Treeatlan data base (Oliveira-

Filho, 2009) (Figure 2). A set of temporal images from the Landsat TM sensor

were used as auxiliary data. The Landsat images from each location were

acquired in the dry and wet seasons in order to identify fragments of deciduous

forests through visual interpretation.

The NDVI time series were derived from the MOD13 product, which

has a spatial resolution of 250m, and 16-day compositing period.

Figure 2 – Locations of Deciduous Forests

The original images were preprocessed using the MODIS Reproduction

Tool (MRT). The data set was sampled to 23 values per year, approximately

two images per month. This dataset included the years of 2003, 2004, and 2005.

2.5 MODIS compositing Methods:

The presence of factors apart from land surface characteristics such as

cloud contamination, atmospheric variability, and bi-directional reflectance

affect the stability of the satellite derived NDVI. Thus compositing methods

have been developed to eliminate these effects. The compositing Method for the

AVHRR NDVI data source is the MVC (Maximum value composite), which

selects the maximum NDVI value on a per pixel basis over a set of compositing

period (Wang et al., 2004).

The MODIS compositing method operates on a per-pixel basis and

relies on multiple observations over a 16-day period to generate a composite

Vegetation Indice. Due to sensor orbit overlap and multiple observations, a

maximum of 64 observations may be collected in a 16 compositing period.

Once all the 16 days of observation are collected, the MODIS VI algorithm

applies a filter to the data based on quality, cloud, and viewing geometry. Only

the higher quality cloud free, filtered data are retained for compositing. (Huete et

al. 2002).

At regional and larger scales, variations in community composition,

micro and regional climate regional, climate regimes, soils, and land

management result in complex spatio-temporal variation in phenology. Further,

some vegetation types exhibit multiple modes of growth and senescence within a

single annual cycle. Therefore compositing methods need to be sufficiently

flexible to allow for this type of variability. (Zhang et al., 2003).

2.6 Fourier Transform

The Fourier Transform has been traditionally used to solve differential

equations in Mathematics and Physics. Its main objective is to approximate a

function in the time domain by a linear combination of harmonics (sinusoids)

(Morettin, 2006). The most basic property of the sinusoids that makes them

suitable for the analysis of time series is their simple behavior under a change in

time scale (Bloomfield, 1976).

Fourier analysis have been used for denoising and curve fitting in

MODIS vegetation index data sets (Wang et al., 2004; Yu et al., 2004; Bruce et

al., 2006; Colditz et al., 2007). If the original time series is discrete rather than

continuous, the Discrete Fourier transform (DFT), which requires regular

spacing on samples within the temporal domains, should be applied (Wang et

al., 2004). Eq 1 depict the DFT: The original signal, x[n], which has N samples.

The transformation will create two vectors containing N/2 values each where

ReX is the Real part vector and ImX is the imaginary part vector of the

transformation, k is the index of these transformation vectors. Eq 2 depicts the

inverse fourier transform or synthesis equation in discrete time, where the

original signal [x] can be completely resynthesized from the ImX and the ReX

vectors.



−

) /Nik cos(2 ] [ix ReX[k]

(1)



−

) /Nik sin(2 ] [ix ImX[k]

(2)



) /Nik sin(2 ][k X Im) /Nik cos(2 ][k X Re][

ππ

(3)

Equations 3, 4 and 5 – Fourrier analys equation in discrete time, Equation (5) sysnthesis

equation in frequency domain – Source Smith (1998)

The algorithm chosen to implement the Discrete Fourier Transform was

the HANTS algorithm (Harmonic Analysis of Time Series) (Verhuef, 1996;

Roerink et al., 2000). The algorithm was developed to deal with time series of

irregularly spaced observations and to identify and remove cloud contaminated

observations. Since the NDVI time series of this study were acquired through

compositing, the pixels have different acquiring dates that lead unequal time

spacing.

HANTS considers only the most significant frequencies expected to be

present in the time profiles (determined, for instance, from a preceding FFT

analysis), and applies a least squares curve fitting procedure based on harmonic

components (sines and cosines) (Verhoef, 1996; Roerink et al., 2000). For each

frequency, the amplitude and phase of the cosine function is determined during

an iterative procedure. Input data points that have a large positive or negative

deviation from the current curve are removed by assigning a weight of zero to

them. After recalculation of the coefficients on the basis of the remaining points,

the procedure is repeated until a predefined maximum error is achieved or the

number of remaining points has become too small. (Roerink, et. al. 2000).

The algorithm runs starting in the upper left block with the original

NDVI time series which are used as input in the FFT and the frequencies that

contain biophysical features (usually mean, annual and half-year signal) are

selected from the Fourier spectrum. The inverse transforms the spectrum back

into a filtered NDVI time-series afterwards. A comparison between the filtered

NDVI time-series and the original NDVI time-series is then made and this is

accomplished by subtracting the corresponding values of each time series. Each

value below a user defined threshold is removed from the time series and

considered cloudy, and then are replaced by values in the filtered NDVI time-

series. (Figure 3). Replacing values in the NDVI time-series makes the average

of the entire profile larger making a next iteration necessary, so the NDVI time-

series is searched again for possible cloud contaminated NDVI observations.

This process continues until no new points are found. Many different

phenological indicators have been defined in various satellite-based studies. The

advantage of the HANTS algorithm is that the output consists of a completely

smoothed NDVI profile which is convenient for calculating derivatives. (De

Wit, 2005) The calculations of derivates are important to estimate the start of

growing season and senescence (Sakamoto et. al 2005).

The version of HANTS used was implemented in IDL by De Wit (2005)

and is under the GNU General Public License.

NDVI

Time-series

Apply

FFT

Select

harmo-

nics

Apply

iFFT

Filtered

Time-series

Original

NDVI

Time-series

Com-

pare

Any

NDVI

points below

user-defined threshold ?

Up-

date

yes

HANTS Filtered

NDVI Time-series

Figure 3 - The HANTS algorithm - Source (De Wit, 2005)

Among the resulting files, the algorithm outputs a FFT file which has a

complex number pair, that is the Fourier transform of each pixel location

regarding its NDVI time series.

2.7 Calculation of amplitude and phase

Harmonic analysis can be used aiming at reducing the dimensionality of

the data. Another advantage is that each pixel is treated individually, being

independent from the rest of the image. It is also possible to choose the period of

analysis relating to the frequency of the studied phenomenon, thus this technique

is appropriate to deal with noise originated from cloud contamination in the time

series and from noise resulting from pre-processing that is not periodic. The

magnitude and phase of a waveform can be calculated from the complex number

resulting from the FFT. The magnitude corresponds to half of wave’s peak

value, and the phase corresponds to the shift from the origin to the wave’s peak

value from 0 to  (eq. 6 and 7) (Lacruz & Santos, 2007).The amplitude/phase

vector corresponds to the polar form of the DFT (Smith, 1998). The output of

the HANTS algorithm contain the complex form for each harmonic (eq. 3, 4

and5) and the mean value of the time series (this value was also used for

discussion and comparison among the different areas).

Since the content in the original FFT prior to the transformation to the

polar form is not intuitive (Smith, 1998), equations 6 and 7 were used to

compute the amplitude and phase of each harmonic. Coldiz et al. (2007) suggest

that only the amplitude and phase of the first three harmonics depict biophysical

parameters. Some authors, such as Yu et al. (2004), state that forest

classification can be carried out in the amplitude/phase space. Previous work by

Oliveira et al. (2009) suggests, however, that information is lost in the

dimensionally reduction and efficient forest classification is not possible in a

rich and complex environment such as the Cerrado.

Bj² Aj² Cj +=

(4)

tan=

(5)

Equations 6 and 7 - Calculation of amplitude (equation 1) and phase (equation

2) for to the polar form of the DFT

2.8 Phase Statistics of NDVI of Deciduous Forests

After the calculation of amplitude and phase of each harmonic of the

original NDVI images (which uses equation 4 and 5) of different locations, a

subset of the original images was generated. Equation 4 calculated the amplitude

of first harmonic. The phase was calculated using eq 5. A subset containing

only the pixels corresponding to the occurrences of Deciduous Forests. Statistics

was computed to the phase of the first harmonic so that annual shifts of

deciduous forests of different geographic locations cold be quantified. Each

phase value ranges from -

, where 2

corresponds to a full year cycle.

The phase values were multiplied by 182.5, which correspond to half year, in

order to calculate annual shifts in days.

3 – Results

Results in table 1 explicit differences in the phase of the annual

frequency of the NDVI value of the deciduous Forests of Minas Gerais. These

results can be very useful for future vegetation classification and to quantify the

geographic differences among apparently similar fragments of this

phytophysiognomy.

The largest time shift was observed for the Triângulo Mineiro Region

(Western Minas Gerais) which is on average 13.45 days ahead in the annual

cycle when compared to the Northern region. This difference could be explained

by the fact that the Deciduous Vegetation in this region is mixed with other

phytophysiognomies such as the Savanna resembled “Cerradão” and other

formations. The mean value of this phytophysiognomy´s time series does not

differ substantially from the others (Table 1). These similarities in the mean

NVDI time series value confirm that deciduous forests do not have discrepancies

in their amplitude value suggesting that the analyzed forest fragments are not

mixed with other vegetation that have higher mean NDVI value such as the

semi-deciduous forests.

Table 1 – statistics of harmonic analysis of the four different study areas

Previous work from Sakamoto et al. (2005) rely on the use of derivates

and wavelet transforms to obtain the days of harvest and plantation of paddy rice

in Japan with the use of MODIS NDVI images. Changes in cropping system,

management, and climate make the times-series collected over agricultural areas

closer to non stationary signals, which are better handled by the wavelet

transform. In the case of native forests that exhibit a stationary behavior, our

FFT approach is most suitable.

The proximity in results regarding mean value of the phase in the North

and the Northwest areas (table 1) can be partly explained by the geographical

proximity of the areas, thus reinforcing that there is a shift in the annual cycle of

Region Fundamental

harmonic’s phase

mean value

Phase value

expressed in days

shifted value

Time

series

mean

value

North East 1.04325 60.6 0.687018

North 1.035954 60,2 0.641661

South 1.141607 66.2 0.735416

West (Triângulo Mineiro) 1.267896 73.65 0.685943

the deciduous forests due to geographical differences. The Southern area also

has a 6 days shift in the average phase value of the annual NDVI frequency and

thus also reinforces our hypothesis.

4-Conclusions

This research suggests that there is an annual shift in the phenological

curve of the Deciduous forests of Minas Gerais that are geographically distant,

however these differences are not great in value. Different regions demonstrate

different annual shifts in the time profile of their deciduous forests of about half

month. However, the spatial resolution of the MODIS sensor limit its application

resulting in few pixels to calculate statistics on small forest fragments such as

the West region of Triângulo Mineiro. The combined use of MODIS NDVI time

series and higher spatial resolution sensors, ground truth data and Geostatistics

might improve the discrimination of deciduous forests in Minas Gerais.

5-References

BLOOMFIELD, P. Fourier analysis of time series. New York: J.Wiley, 1976.

257p.

BRUCE, L.M.; MATHUR, M.; BYRD JUNIOR, J.D. Denoising and wavelet-

based feature extraction of MODIS multi-temporal vegetation signatures.

Geoscience & Remote Sensing, v.43, p.170-180, 2006.

CARVALHO, L.M.T. Mapping and monitoring forest remnants: a multi-

scale analysis of spatio-temporal data. 2001. 140p. Thesis (Doctor in Resources

Conservation and Production Ecology)-Wageningen University, Wageningen.

CARVALHO, L.M.T. de; CLEVERS, J.G.P.W.; SKIDMORE, A.K.; JONG,

S.M. de. Selection of imagery data and classifiers for mapping Brazilian

semideciduous Atlantic forests. Internacional Journal of Applied Earth

Observation and Geoinformation, Enschede, v.5, p.173-186, 2004.

CARVALHO, L.M.T.; SCOLFORO, J.R. Inventário florestal de Minas

Gerais: monitoramento da flora nativa 2005-2007. Lavras: UFLA, 2008. 312p.

CARRÃO, H.; GONÇALVES, P.; CAETANO, M. Contribution of

multiespectral and multitemporal information from MODIS images to land

cover classification. Remote Sensing of Environment, New York, v.112, n.3,

p.986-997, 2007.

COLDITZ, R.R.; GESSNER, U.; CONRAD, C.; ZYL, D. van; MALHERBE, J.;

NEWBY, T.; LANDMANN, T.; SCHMIDT, M.; DECH, S. Dynamics of

MODIS time series for ecological applications in southern África. In:

INTERNATIONAL WORKSHOP ON THE ANALYSIS OF

MULTITEMPORAL REMOTE SENSING IMAGES, 14., 2007, Leuven,

Belgium. Proceedings... Leuven: Multitemp, 2007. p.18-20.

GALFORD, G.L.; MUSTARD, J.F.; MELILLO, J.; GENDRIN, A.; CERRI,

C.C.; CERRI, C.E.P. Wavelet analysis of MODIS time series to detect

expansion and intensification of row-crop agriculture in Brazil. Remote Sensing

of the Environment, New York, v.112, n.2, p.576-587, 2007.

HUETE, A.; DIDAN, K.; MIURA, T.; RODRIGUEZ, E.P.; GAO, X.;

FERREIRA, L.G. Overview of the radiometric and biophysical performance of

the MODIS vegetation indices. Remote Sensing of Environment, New York,

v.83, n.1, p.195-213, Nov. 2002.

JENSEN, J.R. Remote sensing of the enviroment: an earth resource

perspective, retune hall series in geographic information science. New Jersey:

Upper SaddleRiver, 2000. 544p.

LACRUZ, M.S.P.; SANTOS, J.R. Monitoriamento da paisagem de unidades de

conservação. In: RUDORFF, B.F.T.; SHIMABUKURO, Y.E.; CEBALLOS,

J.C. (Org.). O sensor MODIS e suas aplicações no Brasil. São José dos

Campos: Parêntese, 2007. cap.13, p.173-183.

MORETTIN, P.A.; TOLOI, C.M. Análise de séries temporais. 2.ed. São Paulo:

E.Blucher, 2006. 433p.

OLIVEIRA-FILHO, A.T. TreeAtlan 1.0: flora arbórea da Mata Atlântica e

domineos adjacentes: um banco de dados envolvendo geografia diversidade e

conservação. Available at: <http://www. treeatlan.dcf.ufla.br/>. Access in: 10

Jan. 2009.

OLIVEIRA-FILHO, A.T.; CURI, N.; VILELA, E.A.; CARVALHO, D.A.

Effects of canopy gaps, topography and soils on the distribution of woody

species in a central Brazilian deciduous dry forest. Biotropica, Washington,

v.30, n.3, p.362-375, 1998.

OLIVEIRA-FILHO, A.T.; RATTER, J.A. The Cerrados of Brazil: ecology and

natural history of a neotropical savannah. New York: Columbia University,

2002. 121p.

OLIVEIRA- FILHO, A.T.; SCOLFORO, J.R.; CARVALHO, L.M.T.

Mapeamento e inventário da flora nativa e dos reflorestamentos de

Minas Gerais. Lavras: UFLA, 2008. 46p.

OLIVEIRA, L.T. Fusão de imagens de sensoriamento remoto e mineração de

dados geográficos para mapear as fitofisionomias do bioma Cerrado. 2004.

131p. Dissertação (Mestrado em Manejo Ambiental)-Universidade Federal de

Lavras, Lavras.

OLIVEIRA, T.C.A.; OLIVEIRA, L.T.; CARVALHO, L.M.T.;

MARTINHAGO, A.Z.; FAUSTO, W.; LIMA, L.P.Z. Separabilities of forest

types in amplitude-phase space of multi-temporal MODIS NDVI. In: SIMÓSIO

BRASILEIRO DE SENSORIAMENTO REMOTO, 16., 2009, Natal, RN.

Anais... Natal, 2009. No prelo.

ROERINK, G.J.; MENENTI, M.; VERHOEF, W. Reconstructing Cloudfree

NDVI composites using Fourier analysis of time series. International Journal

of Remote Sensing, Basingstoke, v.21, n.9, p.1911-1917, 2000.

SAKAMOTO, T.; YOKOZAWA, M.; TORITANI, H.; SHIBAYAMA, M.;

ISHITSUKA, N.; OHNO, H. A crop phenology detection method using time-

series modis data. Remote Sensing of Environment, New York, v.96, p.366-

374, 2005.

SMITH, S.W. The scientist and engineer's guide to digital signal processing.

San Diego: California Technical, 1998. 626p.

VERHOEF, W. Application of harmonic analysis of NDVI time series

(HANTS). In: AZZALI, S.; MENENTI, M. (Ed.). Fourier analysis of temporal

NDVI in the Southern African and American continents. Wageningen: DLO

Winand Staring Centre, 1996. p.19-24. (Report 108).

WANG, Q.; TENHUNEM, J.; DINH, N.Q.; REICHSTEIN, M.; VESALA, T.;

KERONEN, P. Similarities in ground and satellite-based NDVI time series and

their relationship to physiological activity of a Scots pine forest in Finland.

Remote Sensing of Environment, New York, v.93, n.2, p.225-237, 2004.

WIT, A. de; SU, B. Deriving phenological indicators from spot-vgt data

using the Hants algorithm. Wageningen: Centre for Geo-Information, 2005.

Available at: <http://www.vgt.vito.be/vgtapen/pages/fullpapers/Dewit_full.pdf>.

Access in: 15 dez. 2008.

YU, X.; ZHUANG, D.; CHEN, H.; HOU, X. Forest classification based on

MODIS time series and vegetation phenology. International Geoscience and

Remote Sensing Symposium, Toulouse, v.4, p.2369-2372, 2004.

ZHANG, X.; FRIEDL, M.A.; SCHAAF, C.B. Monitoring vegetation phenology

using modis. Remote Sensing of Environment, v.84, p.471-475, 2003.Oliveira,

T. C. A. (2009) , Oliveira L. T.,Carvalho L.M. T. , , Z. Martinhago, A. Z. ,

Fausto W. Acerbi Júnior, Lima, L. P. Z., Separabilities of Forest Types in

Amplitude-phase Space of multi-temporal MODIS NDVI Anais do IVX

Simósio Brasileiro de Sensoriamento Remoto, no prelo

Sakamoto, T. ; Yokozawa, M. ; Toritani, H. ; Shibayama M. ; Ishitsuka N.;

Ohno H. (2005) A Crop Phenology Detection Method Using Time-Series Modis

Data; Remote Sensing of Environment 96 366 – 374

Mather, P. M. Computer Processing Of Remotely-Sensed Images: An

Introduction. 2. ed. Nottingham, UK: Johb Wiley, 1999, 292 p.

Chapter 02

Mapping Deciduous forests using time series of filtered MODIS NDVI and

Neural Networks

Thomaz C. de A. Oliveira

, Luis M. T. de Carvalho

Luciano T. de Oliveira

, Adriana Z. Martinhago

Fausto W. Acerbi Júnior

1,2,3,4,5

Departamento de Ciências Florestais, Universidade Federal de Lavras

(UFLA).

Caixa Postal 3.037 – 37200-000 – Lavras – MG – Brazil

e-mail:

thomazchaves@gmail.com,

[email protected].br,

dricazm@gmail.com,

[email protected],

fausto@ufla.br

(Prepared According To Cerne)

Abstract: Multi-temporal images are now of standard use in remote sensing

of vegetation during monitoring and classification. (Jensen 2000) Temporal

vegetation signatures (i. e., vegetation indices as functions of time) generated,

poses many challenges, primarily due to signal to noise-related issues Bruce et

al. (2006). This study investigates which methods best generate smoothed

curves of vegetation signatures on MODIS NDVI time series. The filtering

techniques compared were the HANTS algorithm which is based on Fourier

analyses and Wavelet temporal algorithm which uses the wavelet analysis to

generate the smoothed curves. The study was conducted in four different

regions of the Minas Gerais State. The smoothed data were used as input data

vectors for vegetation classification by means of Artificial neural networks. A

comparison of the results was ultimately discussed in this work where results

were encouraging, with no better of the two filtering techniques.

Resumo: Imagens multi-temporais de uso essencial no Sensoriamento Remoto,

para o monitoramento e classificação da vegegetação. Jensen (2000)

Assinaturas temporais da vegetação possuem muitos desafios na sua utilização

devido a elevada relação sinal/ruído Bruce et al. (2006). Este estudo investiga

é o melhor entre dois métodos para se gerar assinaturas temporais suavizadas

de vegetação NDVI sendo essas originadas do sensor MODIS. As técnicas de

filtragem utilizadas foram o algoritmo baseado em Fourier HANTS e algoritmo

Wavelet Temporal que utiliza análise Wavelet.. Os estudos foi conduzido em 4

diferentes conjuntos de dados, correspondente á áreas separadas

geograficamente no estado de Minas Gerais. As séries temporais foram usadas

como entradas de dados para classificação da vegetação através de métodos

computacionais automatizados. Essa mesma foi feita através das redes neurais

artificiais. O resultado dessa classificação foi discutido posteriormente nesse

trabalho com resultados promissores onde não houve melhor entre os métodos

de filtragem comparados

Keywords: remote sensing, signal processing, time series, wavelets analysis,

NDVI, MODIS, Fourier.

Palavras Chave: Sensoriamento Remoto, processamento de sinais, análise

wavelets, NDVI, MODIS, Fourier

1-INTRODUCTION

Mapping land cover by means of remotely sensed data has been a

research of growing interest in the past decades. Its peculiarities and state of art

of computer aided methods and studies go beyond conventional cartographical

tools. (Carvalho, 2001). The advances of computer science, engineering, and

all sciences surrounding remote sensing, continue to present new technologies to

map land cover.

Some objects on the Earth´s surface reflect the electro-magnetic energy

in the same way when sensed with a multi-spectral scanner, in addition ‘objects’

reflectance may vary according to growth stage, phenology, humidity,

atmospheric transparency, illumination conditions, etc. These drawbacks led to

a search for alternative attributes to enable the discrimination of land cover

classes with similar reflectance behavior. (Carvalho et al., 2004)

These attributes, especially temporal information, are very useful for

characterizing deciduous forests in the Cerrado biome, due to their pronounced

dynamics. This can be noticed in the official forest map of Minas Gerias state,

carried out by Carvalho (2008), which does not capture deciduous forests

fragments present in the region of the Triângulo Mineiro, due to the time of

acquisition of the available images. For example, there is a 25ha fragment of dry

Deciduous Forests on the margins of Rio Paranaíba, in fazenda Bonanza,

municipality of Santa Vitória (Oliveira-Filho et al., 1998), which was

misclassified as semideciduous forest. Research however, suggests that

remotely sensed time series data could possibly improve misclassification and

the accuracy of mapping deciduous forests (Oliveira, 2004).

The main objective of this work is to develop an efficient methodology

to map the deciduous forests present in the Cerrado biome based on the use of

temporal attributes. Our research seeks to find whether MODIS filtered NDVI

(Normalized Difference Vegetation Index) time series can generate accurate

mapping to deciduous forests in different regions. This study also has the goal

of testing which of two filtering techniques produces the best map. For

generalization purposes, the process of filtering and mapping were conducted in

four geographically different regions, thus resulting in four different data sets for

comparison purposes. The filtering techniques utilized were wavelet filtering

and the Discrete Fourier Analysis using the HANTS algorithm. The filtered

time series were used as input vectors for each pixel location on artificial neural

networks, both for training and classification, generating vegetation

classification maps.

2-METHODS

2.1 Analysis of Vegetation Temporal Signatures

According to Zhang et al. (2003), field-based ecological studies have

demonstrated that vegetation phenology tends to follow relatively well defined

temporal patterns. For example, in deciduous vegetation and many crops, leaf

emergence tends to be followed by a period of rapid growth, followed by a

relatively stable period of maximum leaf area. Different types of vegetation have

different temporal growth patterns (i.e., different growth and senescence rates)

(Bruce et al., 2006). Vegetation dynamics indicate important short and long-

term ecological process. Continuous temporal observations of land surface

parameters using remote sensing reveal seasonal and inter-annual developments.

Vegetation indices have been extensively applied to characterize the state and

dynamics of vegetation, in particular multiple NDVI (Normalized Difference

Vegetation Index) datasets of the Advanced Very High Resolution Radiometer

(AVHRR) instrument used during the last 25 years. (Jensen, 2000; Coldiz et al.,

2007).

Different vegetation types exhibit distinctive seasonal patterns on NDVI

variation (Yu et al., 2004). A vegetation index should maximize sensitivity to plant

biophysical parameters; normalize or model external factors such as sun angle,

viewing angle, and the atmosphere for consistent spatial and temporal comparisons;

normalize internal effects such as canopy background variations; and couple with

measurable biophysical parameters such as biomass, LAI, or APAR (Jensen, 2000).

Spatial and temporal variability in vegetation indices arise from several

vegetation related properties, including LAI, canopy structure/architecture,

species composition, land cover type, leaf optics, canopy crown cover,

understory vegetation, and green leaf biomass (Huete et al. 2002).

2.2 The MODIS Sensor

The EOS (Earth Observing System), leaded by NASA, has the objective

of studying the Earth’s global changes, its processes and to promote its

continuous observation. Their sensors were designed to operate for a long

period of time. TERRA was the name given to the first platform launched by the

EOS, which marked the development of remote sensing scientific methods by

incorporating various sensors that collect different types of data. The MODIS

(Moderate-resolution Imaging Spectroradiometer) is the most important sensor

aboard the TERRA platform. Its concept has its origin in various predecessors

by which the most important is the AVHRR (Advanced Very High Resolution

Radiometer) aboard the NOAA (National Oceanic and Atmospheric

Administration), from 1978 until 1986. (Soares et al., 2007)

The AVHRR sensor was originally designed for meteorological

applications, and has only two spectral bands (red and near-infrared) that can be

used to generate the spectral indices of vegetation. The new generation MODIS

sensor has a number of advantages over AVHRR, including more spectral bands

that can be used for vegetation analysis (Yu et al., 2004).

MODIS Vegetation Indices (VI) products are appropriate for vegetation

dynamics studies and characterization. MODIS-VI are found to be sensitive to

multi-temporal (seasonal) vegetation variations and to be correlated with LAI

across a range of canopy structure types, species and life forms, land cover

variations. The MODIS NVDI demonstrates an appropiate dynamic range and

sensitivity for monitoring and assessing spatial and temporal variations in

vegetation amount and condition. The seasonal profiles outperform in sensitivity

and fidelity the equivalent AVHRR-NDVI profiles, particularly in atmosphere

with water vapor contents. (Huete et al., 2002)

2.3 Study site

The widespread occurrence of deciduous forests in the state of Minas

Gerais has led to the choice of four different study areas (Figure 1) according to

the locations of deciduous forest fragments extracted from the Treeatlan data

base. (Oliveira-Filho, 2009)

2.4 Data acquisition

The NDVI time series was derived from the MOD13 product, which has

a spatial resolution of 250m, and has a 16-day compositing period. This product

is freely available for download from the MODIS website via FTP protocol.

The original data was reprojected using MRT(MODIS Reproduction Tool). A

set of temporal images from the Landsat TM sensor were also used as auxiliary

data. These images were collected from summer and winter dates to each area

and were used so that the exact locations of the deciduous forests fragments and

other types of vegetation were known in advance. The data set has 23 images per

year and included the years of 2003, 2004 and 2005.

Figure 1 - The four different data subsets

2.5 Signal denoising

In order to extract pertinent features from time signatures for potential

target applications, the signals must first be denoised. Authors have investigated

automated methods for denoising, including straightforward methods such as

median filters and moving-average filtering, as well as more advanced methods

such as wavelet denoising (Bruce et al., 2006). Curve fitting parameterization

using logistic functions have also succeeded in generating time signatures of

MODIS VI (Zahng et al., 2003).

2.6 Fourier Analysis HANTS

The Fourier Analyses or Harmonic analyses have been used traditionally

to solve differential and partial equations in the fields of mathematics and

physics. Its main objective is to approximate a function in the time domain by a

linear combination of harmonics (sinusoids) (Morettin & Toloi, 2006).

The most basic property of the sinusoids that makes them suitable for

the analysis of time series is their simple behavior under a change in time scale

(Bloomfield, 1976).

Fourier analysis have been traditionally used for denoising and curve

fitting in MODIS VI data sets (Wang et al., 2004; Yu et al., 2004; Bruce et al.,

2006; Colditz et al., 2007). If the original data is discrete rather than continuous,

the discrete Fourier transform (DFT), which requires regular spacing on samples

within the temporal domains, should be applied (Wang et al., 2004).

There is a drawback in this approach, since the NDVI images are

composite images of different dates. The pixels have different acquiring dates

that lead unequal time spacing. However, the HANTS (Harmonic Analysis of

Time Series) algorithm was developed to deal with time series of irregularly

spaced observations and to identify and remove cloud contaminated

observations (Verhuef, 1996; Roerink et al., 2000).

The HANTS algorithm considers only the most significantcant

frequencies expected to be present in the time profiles (determined, for instance,

from a preceding FFT analysis), and applies a least squares curve fitting

procedure based on harmonic components (sines and cosines) (Verhoef, 1996;

Roerink et al., 2000). For each frequency the amplitude and phase of the cosine

function is determined during an iterative procedure. Input data points that have

a large positive or negative deviation from the current curve are removed by

assigning a weight of zero to them. After recalculation of the coefficients on the

basis of the remaining points, the procedure is repeated until the maximum error

is acceptable or the number of remaining points has become too small. (Roerink

et al., 2000).

Many different phenological indicators have been defined in various

satellite-based studies. The advantage of the HANTS algorithm is that the output

consists of a completely smoothed NDVI profile which is convenient for

calculating derivatives. (De Wit & Su, 2005) The calculations of derivates are

very important for the estimation the start of growing season and senescence

dates (Sakamoto et al., 2005).

The version of HANTS used was implemented in IDL by (DE WIT,

2005) and is under the GNU General Public License.

The data set of temporal NDVI images, which contained a series of 23

samples per year was input to the HANTS algorithm. The output is a similar

time series but smoothed and containing only the annual, 6 months and 3 months

frequencies of the signal. The resulting different data sets were used as input

data sets for image classification.

2.7 Wavelet transform

Fourier series are ideal for analyzing periodic signals, since harmonics

modes used in the expansions are themselves periodic. In contrast, the Fourier

integral transform is a far less natural tool because it uses periodic functions to

expand nonperiodic signals. Two possible substitutes are the windowed Fourier

transform (WFT) and the wavelet transform. The windowed Fourier transform

can, however, be an inefficient tool to analyze regular time behavior that is

either very rapid or very slow relative to the size of the analyzing window. The

Wavelet transform solves both problems by replacing modulation with scaling to

achieve frequency localization. The WFT might also be an inefficient tool when

very short time intervals are of interest. On the other hand, a similar situation

occurs when very long and smooth features of the signal are to be reproduced by

the WFT. (Kaiser, 1994).

Different from the infinite sinusoidal waves of the Fourier transform, a

wavelet is a small wave localized in time or space. Since a wavelet has compact

support, which means that its value becomes 0 outside a certain interval of time,

the time components of time-series can be maintained during the wavelet

transformation. (Sakamoto et al. 2005)

Previous work reveal that the wavelet transform is a powerful tool for

denoising data sets and for curve fitting procedures in NDVI time series

(Sakamoto et al., 2005; Bruce et al., 2006; Galford et al., 2007).

For the present work, we used the methodology proposed by (Carvalho,

2001). In remote sensing outliers caused by clouds and shadows (noise) appear

as peaks with narrow bandwidth in the temporal spectrum. They appear similar

in the spatial domain, but with variable bandwidth. If we consider the presence

of clouds and shadows as signal response against a “noisy” background, a

framework for their detection can be based on noise modeling in transformed

space. The discrete wavelet transform was implemented with the ‘à trous’

algorithm with a linear spline as the wavelet prototype. It produces a vector of

wavelet coefficients d at each scale j, with j=0,…,J. The original function f(t)

was then expressed as the sum of all wavelets scales and the smoothed version

. The input signal was decomposed using one scale, two scales and three

scales. The resulting different data sets were used as inputs for image

classification, described in the following section.

2.8 Image Classification

The main objective of this work is to compare different filtering

techniques and their output vegetation signature for time series of NDVI. One

way to accomplish this is to use smoothed time series as input vectors to

automated image classification.

For Moreira (2003) automatic image identification and classification can

be understood as the analyses and the manipulation of images through

computational techniques, with the goal of extracting information regarding an

object of the real world.

2.9 Artificial Neural Networks

Humans and other animals process information with neural networks.

These are formed from trillions of neurons (nerve cells) exchanging brief

electrical pulses called action potentials. Computer algorithms that mimic these

biological structures are formally called artificial neural networks to distinguish

them from the squishy things inside of animals (Smith, 1998). These biological

inspired models are extremely efficient when the pattern of classification is not a

simple and trivial one. Theses networks have shown to be helpful in the

resolution of problems of practical scope. Problems such as voice recognition,

optical character recognition, medical diagnosis and other practical scope

problems are by no means complex problems to the human brain and sensor as

they are for a computer to resolve.

Even though, some researchers do not recognize the artificial neural

networks as being the general natural solution surrounding the problems of

recognizing patterns on processed signals, it can be noticed that a well trained

network is capable of classifying highly complex data. The use of artificial

neural networks in pattern recognition and classification has grown in the last

years in the field of remote sensing (Kanellopolous, 1997).

This work proceeded with 2 filtered data sets per region, these data sets

included one HANTS filtered time series and one Wavelets filtered time series.

These data sets were input into a neural network with the following

characteristics: sigmoidal activation function, 0.01 learning rate, momentum

factor of 0.5, sigmoid constant of 1.0, 14 hidden layers, with 69 neurons per

layer. For training the network, 10000 iterations were used, with RMS error of

0.0001. These parameters were extracted from literature based on standard

applications of neural to remote sensing image classification.

3-RESULTS AND DISCUSSION

Classification results as shown in table 1 confirm that no time series

filtering technique necessarily produces a more accurate classified map. In some

cases the classified maps produced from HANTS filtered time series generated

more accurate results. In others cases it produced less accurate results. The

kappa coefficient for the classification results can be either classified as

substantial or almost perfect (Landis & Kock, 1977). Different from our

findings, previous work carried out by Burce et al. (2006), which have also used

filtered time series from Wavelet and Fourier transforms for image

classification, showed that the former produced more accurate results. This can

be partly explained by the fact that the HANTS algorithm have some

enhancements over traditional Fourier based algorithms which was present in the

cited work.

Table 1 – classification results

The rows of figure 2 show the classification results in the four study

areas. The results of Wavelet filtering time series, used as input for classification

are in the right hand column. In the left hand column, the HANTS filtered time

series as input to the neural network are illustrated. In the middle column we

have the official forest map of Minas Gerais, carried out by Carvalho & Scolforo

(2008), with a 30m spatial resolution. The proposed methods captured the

general characteristics of vegetation of each area. In some cases such as the

North East area, the classification results resemble the general “shape” of forest

fragments. Both Northern area data sets have similar patterns when compared to

the official state map. The other areas, however, do not show these similarities

with the general shape of forest fragments from the official map. Note that the

spatial resolution has important implications in the map comparisons. The

Northern areas have a more accurate “shaping” of vegetation classification.

Region Kappa Coefficient (wavelet

filtering)

Kappa Coefficient

(HANTS filtering)

North East

0.9480 0.8257

North

0.8476 0.8333

South

0.6355 0.7412

West region (Triângulo

Mineiro)

0.8729 0.9051

North East area classification results

Northern area classification results

Southern area classification results

Western area (Triângulo Mineiro) classification results

Figure 2 – classification results

This work developed an efficient methodology to map the deciduous

forests present in the Cerrado biome using MODIS temporal attributes and

artificial neural networks algorithm. This study concluded that MODIS filtered

NDVI (Normalized Difference Vegetation Index) time series can generate

accurate maps of deciduous forests in different regions.

• The maps generated from both HANTS and wavelet transformation

curve smoothing procedures showed very similar high accuracy,

indicating that any of these procedures can be used to denoise

similar data sets.

• In the Northern areas, the maps generated from temporal features

resemble the general “shape” of forest fragments, having similar

patterns when compared to the official state map

• This methodology was capable of detecting fragments of deciduous

forests in the Triângulo Mineiro region where the official state map

did not.

Future research on this topic could be enhanced by the use of soft

classifiers, such as Fuzzy logic which considers the possibility of one location

belonging to various classes of vegetation. Oliveira-Filho (2009) criticizes the

“rigidness” of the forest classification systems and is also concerned about the

authenticity of trueness regarding these assessments. His work presents a ‘new’

and ‘flexible’ method which could be coupled with the fuzzy logic in the future.

There is, however, the possibility of a chaos injection as predicted by the author,

as the necessity of naming complex structures is ever present and useful.

5. BIBLIOGRAPHICAL REFERENCES

BLOOMFIELD, P. Fourier analysis of time series. New York: J.Wiley, 1976.

257p.

BRUCE, L.M.; MATHUR, M.; BYRD JUNIOR, J.D. Denoising and wavelet-

based feature extraction of MODIS multi-temporal vegetation signatures.

GIScience & Remote Sensing, v.43, p.170-180, 2006.