Automatic tree species recognitionwith quantitative structure modelsMarkku Åkerblom, 1 Pasi Raumonen, 1 Raisa Mäkipää, 2 and Mikko Kaasalainen 1

IntroductionF A novel tree species recognition approach from terrestrial laser scanner (TLS)

data.

F TLS data dimensionality is increased by reconstructing Quantitative StructureModels (QSM) before classification feature computations.

F Tree species recognised from 15 geometrical and topological classificationfeatures derived from QSMs.

F The input of the fully automatic approach is a forest plot level point cloud, theoutput a collection of detailed 3D models with the species information.

Materials and methodsF Included species:

B Silver birch P Scots pine S Norway spruceBetula pendula Roth Pinus sylvestris L. Picea abies [L.] Karsten

F 3 single-species and 2 mixed-species forest plots inFinland with over 1200 trees.

F RIEGL VZ-400 with 0.04 resolution and several scanningpositions.

F Retroreflectors & RiScan Pro for registration.

F Trees extracted from TLS data and reconstructed as QSMsautomatically.

F Classification methods tested: k-nearest neighbours,multinomial regression, and support vector machines.

Classification featuresF Features are robust with respect to tree and branch

extraction errors.

406080

100Stem branch angle

0

0.5

1Crown start height

B P S

00.5

11.5

Cylinder len. / tree vol.

# Stem branch

1 Angle2 Cluster size3 Radius4 Length5 Distance

# Crown

6 Start height7 Height8 Evenness9 Diameter / height

# Tree

10 DBH / tree height11 DBH / tree volume12 DBH / min. tree radius13 Volume below 55 % of height14 Cylinder length / tree volume15 Shedding ratio

B

Results: single-species forest plots

1010trees from three plots

10-fold leave-one-out cross validation by dividing datainto 10 equal sized bins. One-by-one each bin actedas testing data while the other nine formed the trainingdata.

96%maximum accuracy

All classificationmethods performedwell (> 95 %) withsome feature combination. Highest accuracy, 96.9 %with the 4-NN method and 10 features.

30training samples

Training set size was varied in steps from 4 to 150 sam-ples per species. The increase in accuracy after 30 sam-ples was minimal, indicating that as little as 30 trainingsamples per species can be enough.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 150

4080

126Times

into

p

126 top feature combinations were identified: average of all methods ≥ 95 %, orminimum of all methods ≥ 94 %.

Results: mixed-species forest plotsOnly preliminary testing was possible with trees from mixed-species forest plots dueto the lack of comprehensive combined TLS and reference species classification data.76 % maximum accuracy when using only single-species training data, 81 % whentraining data was augmented with spruce trees from mixed-species stands. In thelatter case accuracy was 91 % for spruce trees that had comprehensive training data.

40 60 80 1000

0.2

0.4

0.6

0.8

1

Stem branch angle

Volum

ebelo

w55

%of

height

single-speciesmixed-species

249trees from two plots

81%maximum accuracy

Further informationÅkerblom, M., Raumonen, P., Mäkipää, R., and Kaasalainen, M. (2017).Automatic tree species recognition with quantitative structuremodels. Remote Sensing of Environment, 191:1 – 12.https://authors.elsevier.com/a/1UPdU7qzSbsjg

Åkerblom, M. (2017). Tree species recognition with quantitativestructure models. Youtube.https://www.youtube.com/watch?v=SX6kYeuY0Oo

1 Tampere University of Technology 2 Natural Resources Institute FinlandDepartment of Mathematics (LUKE)

PO Box 553 PO Box 18FI-33101 Tampere FI-01301 Vantaa

Finland Finlandmath.tut.fi/inversegroup www.luke.fi/en