Forest Classification Using High Spectral and Spatial Resolution Data

1
Forest Classification Using High Spectral and
Spatial Resolution Data

• Juho Lumme
• Helsinki University of Technology

2
Introduction

• Forests are important natural resources for
  Finland
• They have financial, biological, social and
  cultural significance
• RS is an efficient tool for forest inventory
• Imaging spectrometer measures spectral and
  spatial properties
• Textural and spectral features are both used
• Data is analysed using separability measures and
  classification

3
Test area

• Area is located in Lammi (Southern Finland)
• Contain
• Lakes
• Rural areas
• Cultivated fields
• Coniferous forests
• Deciduous forests
• Size 50 km long 2 km wide.

4
AISA data

• September 1999
• Six strips
• Pixel size 1 m
• 17 channels (visible light and infrared)
• Field surveys
• Geological Survey of Finland
• Over 250 training areas

5
Preliminary works Feature extraction

• Geometric and radiometric correction (METLA)
• NDVI and image texture were used to screen forest
  areas
• Principal Component Analysis
• Texture features...

6
Texture measures

• Measures were calculated using co-occurrence
  matrix
• Window sizes 25, 35, 45 pixels
• Distances 1, 3, 5 pixels
• 9 texture measures
• Homogeneity, contrast, dissimilarity, mean,
  standard deviation, entropy, angular second
  moment, correlation inverse difference

7
Forest Classes

• birch (seedling)
• birch (young/old)
• spruce (seedling)
• spruce (young)
• spruce (old)
• pine (young)
• pine (old)

8
Separability measures individual features

• Class separation was poor
• Best spectral channel 9
• Best texture measure mean
• Spruce and pine mixed
• Different age forest types mixed

9
Separability measures feature sets

• Both spectral and textural features
• All the classes were separated

10
Image classification

• Bayes Perceptron Neural Network
• Several different feature sets

11
Image classification
1 Bayes classification with principal component
channels 2 Bayes classification with principal
component channels and texture features 3 Bayes
classification with three original image
channels, principal component channels, NDVI and
texture features 4 Neural network
classification with all original image channels
and texture features

• Seedling classes were rather uncertain
• Bayes led to better results

12
Conclusion

• Different age forest stands are hard to
  distinguish
• Both spectral and textural features are useful on
  forest airborne hyperspectral data analysis
• The results of this study are able to generalize
  to the satellite case

13
Grey level co-occurrence matrix

• A GLCM is a two-dimensional histogram of grey
  levels for a pair of pixels which are separated
  by a fixed spatial relationship (PCI, Geomatica)

14
GLCM Texture analysis