Kernel feature cross-correlation for unsupervised quantification of damage from windthrow in forests

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F62156489%3A43410%2F16%3A43910917" target="_blank" >RIV/62156489:43410/16:43910917 - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.5194/isprsarchives-XLI-B7-17-2016" target="_blank" >http://dx.doi.org/10.5194/isprsarchives-XLI-B7-17-2016</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.5194/isprsarchives-XLI-B7-17-2016" target="_blank" >10.5194/isprsarchives-XLI-B7-17-2016</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Kernel feature cross-correlation for unsupervised quantification of damage from windthrow in forests

Popis výsledku v původním jazyce

In this study estimation of tree damage from a windthrow event using feature detection on RGB high resolution imagery is assessed. An accurate quantitative assessment of the damage in terms of volume is important and can be done by ground sampling, which is notably expensive and time-consuming, or by manual interpretation and analyses of aerial images. This latter manual method also requires an expert operator investing time to manually detect damaged trees and apply relation functions between measures and volume which are also error-prone. In the proposed method RGB images with 0.2 m ground sample distance are analysed using an adaptive template matching method. Ten images corresponding to ten separate study areas are tested. A 13 x 13 pixels kernel with a simplified lin ear-feature representation of a cylinder is applied at different rotation angles (from 0o to 170o at 10o steps). The higher values of the normalized cross-correlation (NCC) of all angles are recorded for each pixel for each image. Several features are tested: percentiles (75, 80, 85, 90, 95, 99, max) and sum and number of pixels with NCC above 0.55. Three regression methods are tested, multiple regression (mr), support vector machines (SVM) with linear kernel and random forests. The first two methods gave the best results. The ground-truth was acquired by ground sampling, and total volumes of damaged trees are estimated for each of the 10 areas. Damaged volumes in the ten areas range from c. 1.8 x 102 m3 to c. 1.2 x 104 m3. Regression results show that smv regression method over the sum gives an R-squared of 0.92, a mean of absolute errors (MAE) of 255 m3 and a relative absolute error (RAE) of 34% using leave-one-out cross validation from the 10 observations. These initial results are encouraging and support further investigations on more finely tuned kernel template metrics to define an unsupervised image analysis process to automatically assess forest damage from windthrow.

Název v anglickém jazyce

Kernel feature cross-correlation for unsupervised quantification of damage from windthrow in forests

Popis výsledku anglicky

In this study estimation of tree damage from a windthrow event using feature detection on RGB high resolution imagery is assessed. An accurate quantitative assessment of the damage in terms of volume is important and can be done by ground sampling, which is notably expensive and time-consuming, or by manual interpretation and analyses of aerial images. This latter manual method also requires an expert operator investing time to manually detect damaged trees and apply relation functions between measures and volume which are also error-prone. In the proposed method RGB images with 0.2 m ground sample distance are analysed using an adaptive template matching method. Ten images corresponding to ten separate study areas are tested. A 13 x 13 pixels kernel with a simplified lin ear-feature representation of a cylinder is applied at different rotation angles (from 0o to 170o at 10o steps). The higher values of the normalized cross-correlation (NCC) of all angles are recorded for each pixel for each image. Several features are tested: percentiles (75, 80, 85, 90, 95, 99, max) and sum and number of pixels with NCC above 0.55. Three regression methods are tested, multiple regression (mr), support vector machines (SVM) with linear kernel and random forests. The first two methods gave the best results. The ground-truth was acquired by ground sampling, and total volumes of damaged trees are estimated for each of the 10 areas. Damaged volumes in the ten areas range from c. 1.8 x 102 m3 to c. 1.2 x 104 m3. Regression results show that smv regression method over the sum gives an R-squared of 0.92, a mean of absolute errors (MAE) of 255 m3 and a relative absolute error (RAE) of 34% using leave-one-out cross validation from the 10 observations. These initial results are encouraging and support further investigations on more finely tuned kernel template metrics to define an unsupervised image analysis process to automatically assess forest damage from windthrow.

Druh

D - Stať ve sborníku

CEP obor

GK - Lesnictví

OECD FORD obor

—

Projekt

—

Návaznosti

I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Rok uplatnění

2016

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

XXIII ISPRS Congress: Technical Commission VII

ISBN

—

ISSN

1682-1750

e-ISSN

—

Počet stran výsledku

6

Strana od-do

17-22

Název nakladatele

Copernicus GmbH

Místo vydání

Göttingen

Místo konání akce

Praha

Datum konání akce

12. 7. 2016

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

000393155900003