Flower Species Identification And Coverage Estimation Based On Hyperspectral Remote Sensing Data

1
Flower Species Identification And Coverage
Estimation Based On Hyperspectral Remote Sensing
Data
Gai Yingying1, Fan Wenjie1, Xu Xiru1, Zhang
Yuanzhen2 1. Institute of RS and GIS, Peking
University, Beijing, China 2. China
Meteorological Administration Training Centre,
Beijing, China
Email Address fanwj_at_pku.edu.cn (Fan Wenjie)
2
Outline
1. Preface 2. Data 2.1 Data acquirement 2.2 Data
preprocessing 3. Methodology 3.1 Flower spectral
feature extraction 3.2 Mixed spectra unmixing 4.
Results 5. Discussion
3
Preface

• Causes of grassland degradation
• overgrazing
• excess reclamation

• Superiorities of hyperspectral remote sensing
• provide information at different temporal and
  spatial scales
• high spectral resolution

Monitoring grass species and coverage accurately
using hyperspectral remote sensing data makes a
significant contribution to species diversity
research and sustainable development of grassland
ecosystem. Hyperspectral remote sensing becomes
an important way of monitoring terrestrial
ecosystem.
4
Data

• Data acquirement
• Study area Hulunbeier meadow grassland,
  Hulunbeier City, Inner Mongolia, China.
• Time from July 1st to July 3rd, 2010
• Flower species Serratula centauroides Linn.,
  Clematis hexapetala Pall., Artemisia frigida
  Willd. Sp. Pl., Galium verum Linn., Hemerocallis
  citrina Baroni, Lilium concolor var. pulchellum
  and Lilium pumilum

5
Data

• Data acquirement
• Device ASD FieldSpec-3, with the spectral range
  of 3502500 nm and the spectral resolution of 1
  nm
• Data type spectra of same kind flower canopies,
  spectra of quadrates contained flowers of single
  and multiple species

6
Data

• Data prepocessing
• Wavelet filtering

7
Data

• Data prepocessing
• Comparison of Wavelet filtering and
  Savitzky-Golay filtering

Signals of high frequency were more stable
dealing with wavelet filter than Savitzky-Golay
filter.
8
Methodology

• Flower spectral feature extraction
• Spectral Differential
• --- identify Serratula centauroides Linn. and
  divide other flowers into three sets

1. The spectral derivatives of Serratula
   centauroides Linn. between purple and blue bands
   are below zeros
2. The maximum derivatives of both Clematis
   hexapetala Pall. and Artemisia frigida Willd. Sp.
   Pl. in the range from 500 nm to 600 nm are much
   smaller than others
3. The derivatives of Galium verum Linn. and
   Hemerocallis citrina Baroni reach peaks in
   500-550 nm, while Lilium concolor var. pulchellum
   and Lilium pumilum in 550-600 nm.

9
Methodology

• Flower spectral feature extraction
• Spectral Differential
• --- identify Serratula centauroides Linn. and
  divide other flowers into three sets

(1) (2) (3) (4)
10
Methodology

• Flower spectral feature extraction
• Spectral Reordering
• ---identify Clematis hexapetala Pall. and
  Artemisia frigida Willd. Sp. Pl.

When spectra were reordered based on Clematis
hexapetala Pall., curves of Artemisia frigida
Willd. Sp. Pl. shows different fluctuation. It is
the same the other way round.
11
Methodology

• Flower spectral feature extraction
• Vegetation Index
• ---identify the other two sets Galium verum
  Linn., Hemerocallis citrina Baroni
• Lilium concolor var. pulchellum, Lilium pumilum.

Flower species NDVIs
Galium verum Linn. 0.5119-0.5985
Hemerocallis citrina Baroni 0.2145-0.3224
Flower species ?
Lilium pumilum 2.9407-3.7834
Lilium concolor var. pulchellum 4.1446-9.0796
12
Methodology

• Mixed spectra unmixing
• linear spectral mixture analysis

• Definition
• mixed pixel
• end-member

P --- measured spectra vector N --- number of
end-numbers Ci --- proportion of ei in pixels n
--- error
C --- proportional vector of end-numbers E ---
matrix of end-number vector

• quadrate spectra --- mixed spectra
• flower spectra --- end-member spectra
• range of wave bands--- 400-750 nm

13
Results
Accuracy analysis of flowers identification
Flower species Not-identify error / Incorrect-identify error / Total error/
Serratula centauroides Linn. 8.33 0 8.33
Clematis hexapetala Pall. 0 6.67 6.67
Artemisia frigida Willd. Sp. Pl. 6.67 0 6.67
Galium verum Linn. 5.88 3.03 8.91
Hemerocallis citrina Baroni 0 5.88 5.88
Lilium concolor var. pulchellum 0 0 0
Lilium pumilum 0 0 0
Verification results showed that when the
coverage of flowers was more than 10, the
accuracy of identification methods would be
higher than 90.
14
Results
Accuracy analysis of pixel unmixing method
Flower species Mean error Standard deviation
Serratula centauroides Linn. 0.040 0.065
Clematis hexapetala Pall. 0.042 0.034
Artemisia frigida Willd. Sp. Pl. 0.062 0.032
Galium verum Linn. 0.029 0.073
Hemerocallis citrina Baroni 0.052 0.037
Lilium concolor var. pulchellum 0.021 0.028
Lilium pumilum 0.018 Null
Note There are not enough data for validation of
Lilium pumilum.
Results also showed that the linear unmixing
model was an effective method for estimating the
coverage of flowers in grassland with the mean
error of about 4.
15
Discussion
Discussion
The methods studied in the paper demonstrate
promising application in monitoring some herb
plants during florescence. More flowers will also
be distinguished with high accuracy if
multi-temporal data are available. In our study,
application of field measured hyperspectral data
in vegetation monitoring has been broaden, but
species identification using remote sensing is to
some extent limited by field observation.
Admittedly, what we have observed in this study
is far from complete and it requires further
research.
16
Discussion
Discussion
Grasslands need protection!
17
Thanks !
Email Address fanwj_at_pku.edu.cn (Fan
Wenjie) Institute of RS and GIS, Peking
University, China