Nonparametric and Probabilistic Classification of Agricultural Crops Using Multitemporal Images Smge

1
Nonparametric and Probabilistic Classification
of Agricultural Crops Using Multitemporal
ImagesSmögen Workshop, 21-25 August 2006

• Jun Yu
• Bo Ranneby
• Centre of Biostochastics
• The Swedish University of Agricultural Sciences
• Umeå, Sweden

2
(No Transcript)
3
Input Data

• Field Data
• Block database (marginal part as ground
  truth)
• Block database (for evaluation)
• Satellite Images

4
Crops
25 classes Autumn-sown cereals Spring-sown
cereals Spring-sown oil seed crops Potatoes Gra
ss land on arable land (for hay or silage)
Energy forest (salix) Wood land on pasture
5
Test sites in the County of Dalarna
6
Test sites background GSD topographical map
7
Methodology

• Define the target function
• (in this case, probabilities of correct
  classification)
• Denoise the images
• Remove outliers from reference data
• Calculate the information values in the
  components in the feature vector (e.g. different
  bands)
• Determine a proper metric
• Determine prototypes for the classes
• Run a nonparametric classification so that the
  target function is maximized
• Declare the quality of classification result by
  using probability matrices

8
Classification test site 1
5 scenes
1 scene
9
Classification test site 2
1 scene
5 scenes
10
Probability Matrices
5 scenes
1 scene
11
Probability Matrices at level 1
5 scenes
1 scene
Level 1 C1 arable land C2 pasture and
meadows
12
More quality

• Calculate probabilities for classes at pixel
  level
• Calculate entropy for each pixel

13
Classification test site 1, 5 scenes
14
Probability per class, test site 1, 5 scenes
15
Entropy, five scenes, test site 1
16
Pixelwise probability per class, and entropy
test site 1
Entropy value
17
Entropy, one scene, test site 1
18
Classification test site 2, 5 scenes
19
Probability per class, test site 2, 5 scenes
20
Entropy, five scenes, test site 2
21
Pixelwise probability per class, and entropy
test site 2
Entropy value
22
Entropy, one scene, test site 2
23

• Thank you for your attention!