LINEAR UNMIXING OF MULTIDATE HYPERSPECTRAL IMAGERY FOR CROP YIELD ESTIMATION

1
LINEAR UNMIXING OF MULTIDATE HYPERSPECTRAL
IMAGERY FOR CROP YIELD ESTIMATION

• Bin Luo1, Chenghai Yang2 and Jocelyn Chanussot3
• 1 LIESMARS, Wuhan University, Wuhan, China
• 2 U.S. Department of Agriculture, Weslaco, Texas,
  USA
• 3 Grenoble Institute of Technology, Grenoble,
  France
• IGARSS 2011 24 29 July, 2011 Vancouver,
  Canada

2
Mapping Yield Variation for Precision Agriculture

• Remote sensing imagery has been commonly used for
  estimating crop yield variation
• Vegetation indices (e.g., NDVI)
• With hyperspectral imagery, the number of VIs is
  large
• Spectral unmixing can be used to derive abundance
  images

3
Spectral Mixing

• A pixel can be considered as a mixture of plants
  and soil.
• Spectral unmixing can quantify crop canopy
  fraction within each pixel.
• A crop fraction image is a more direct measure of
  plant abundance than NDVI
• Plant abundance is indicative of crop yield.

4
Objectives and Procedures

• Evaluate unsupervised linear unmixing approaches
  on hyperspectral images for crop yield estimation
• Use multi-date hyperspectral data for improving
  estimation results

VCA (Vertex Component Analysis
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
5
Unmixing of Hyperspectral Images

• Linear mixture model of hyperspectral images
• X MS n
• M unmixing matrix
• S abundance matrix

• VCA (Vertex Component Analysis) to extract
  endmembers
• Red cross
• hyperspecral data X
• Blue circles
• endmembers M
• Abundance S
• Random between 0 1

Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
6
Airborne Hyperspectral Images

• Hyperspectral system
• Spectral range 467932 nm
• Swath width 640 pixels
• Bands 128
• Radiometric 12 bit (04095)
• Pixel size 1 m
• Study site
• Two grain sorghum fields in south Texas
• 13.4 ha and 14.0 ha in size
• Image timing
• Shortly before and after crop reached maximum
  canopy cover
• 18-May-2001 and 29-May-2001

Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
7
Geometric Correction, Rectification Calibration

• Geometric correction
• Reference line approach
• Rectification
• Georeference images to UTM
• with GPS ground control points
• Radiometric calibration
• Three tarps with reflectance of 4, 32, and 48
  were used to convert digital counts to
  reflectance
• 102 bands were used for analysis

Raw
Corrected
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
8
Grain Sorghum Yield Data Collection
Ag Leader PF3000 Yield Monitor
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
9
Yield Data
Crop yield images of the two fields.
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
10
Fusion of Multi-date Unmixing Results
Flow chart of the fusion of the multi-date
unmixing results
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
11
Fusion of Multi-date Unmixing Results

• M18(k) and M29(k) as the abundances of crop
  extracted on the date 18 May 2001 and 29 May 2001
  at the kth pixel
• Evaluation Correlation coefficients

where Y is the yield data
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
12
Fusion of Multi-date Unmixing Results
M18(k) of Field 1
M29(k) of Field 1
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
13
Fusion of Multi-date Unmixing Results
M18(k) of Field 2
M29(k) of Field 2
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
14
Fusion of Multi-date Unmixing Results
Correlation coefficients between the yield data
and the (combined) crop abundances of Field 1
M1 M2 M3 M4
C(Mi, Y) 0.739 0.748 0.780 0.764
Correlation coefficients between the yield data
and the (combined) crop abundances of Field 2
M1 M2 M3 M4
C(Mi, Y) 0.648 0.721 0.735 0.701
Recall that
Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation
15
Conclusions

• Crop abundances obtained by the unsupervised
  linear unmixing are strongly correlated to crop
  yield data.
• The fusion of crop abundances obtained from
  images taken at different dates significantly
  improves the correlation with yield.

Linear Unmixing of Multidate Hyperspectral
Imagery for Crop Yield Estimation