Spectral measurement improvement through optical tree delineation

Abstract

Optical tree delineation algorithms have previously been used for satellite data. In this thesis, they are adapted for use from an Unmanned Aerial Vehicle (UAV). An evaluation of three algorithms (Valley Following, Region Growing and Marker Controlled Watershed) as well as different pre-filtering was done to find the best algorithm and filter combination for a species independent delineation. The Region Growing algorithm gave the best results, and was used to delineate the tree crowns for the Spectral investigation. An SVM classifier was trained to classify the results of the delineation and found that the omission of background points increased the accuracy by 7.7% to 80.3% while reducing processing time by 70 %. Reducing the data amount further by creating a single average over each tree crown, increased classification by a another 1.3% with a processing time reduction of 97 % compared to the full tree stands. The standard deviation of the spectral values within a tree crown where found to carry species information, resulting in an accuracy of 62.5 % or an overall accuracy of 83.3 % when used in combination with the spectral mean tree crown values.