The endemic New Zealand kauri trees (Agathis australis (D.Don) Lindl.) are a key species in New Zealand’s northern indigenous forests. As one of the largest and longest-lived trees in the world, mature kauri are a tourist attraction and have high cultural significance for local Māori. However, the trees are threatened by the deadly kauri dieback disease (Phytophthora agathidicida (PA)). Over the last decade, PA has been detected throughout most of the kauri distribution area. PA is a soil-borne pathogen that enters the trees via the root system and causes collar rot, thereby blocking the transport of water and nutrients to the canopy, eventually killing the tree.

This thesis aims to develop methods based on remote sensing to automatically identify kauri trees and detect stress symptoms in their canopy. It is important to note that canopy stress symptoms are not proof of an infection. The reference data used here include 3165 precisely located crowns from three study sites in the Waitakere Ranges west of Auckland. They cover a representative range of both kauri and associated tree species in different forest ecotypes and stand situations. The selection of kauri crowns includes a range of phenological varieties, such as colour variants, growth stages and stress symptom levels.

The structure of this thesis follows three research questions, which form the basis of three scientific papers. The first paper aims to identify kauri trees with optical remote sensing. A distinct spectral pattern of kauri crowns could be discovered with the use of an airborne AISA Fenix hyperspectral image in the far near-infrared part of the spectrum. The paper presents a method to distinguish kauri with no to medium symptoms from dead and dying tree crowns and other canopy species with no to medium symptoms. High user’s and producer’s accuracies of 94.6% and 94.8% for the class “kauri” were achieved in a Random Forest classification using five spectral indices on five wavelengths (670–1209 nm). The kauri spectra showed a high separability to the spectra of 21 other canopy species and vegetation. However, the distinction between dead and dying trees and other tree species turned out to be more difficult. A minimum crown diameter of 3 m was defined for the 1 m pixel resolution to minimize the effect of mixed pixels. The overall accuracy (OA) for the three target classes could be improved from 91.7% to 93.8% by combining “kauri” and “dead/dying” trees into one class, separately classifying low and high forest stands and a binning to 10 nm bandwidths.

The second paper focuses on an analysis of reflectance patterns for different stress levels and growth stages in kauri crowns. The analysis was again based on hyperspectral images and 1258 manually edited reference crowns of “kauri” and “dead/dying” trees. The field assessment for stress symptoms was complemented with an evaluation of visible canopy symptoms in Red-Green-Blue (RGB) aerial images. An image guideline for stress assessment based on aerial images was developed. A Normalised Difference Vegetation Index (NDVI) in the near-infrared/red spectral range and indices with bands in the near-infrared

and red-edge were identified as the most important band combinations to describe the full range of stress responses. However, pigment-sensitive indices with bands in the green and red spectral ranges are more important for describing first stress symptoms and stress responses in smaller trees with denser foliage. Five indices on six bands in the visible to near-infrared region (450–970 nm) achieved a correlation of 0.93 with a Random Forest regression for the description of five stress symptom levels from non-symptomatic to dead. A stratified approach with individual models for pre-segmented low and high forest stands improved the overall performance. Additional bands in the far near-infrared region improved the root mean square error (RMSE) slightly from 0.43 to 0.42 but not the correlation.

In the third paper, the use of WorldView-2 satellite data (8 multispectral bands, 1.8 m pixel resolution pan-sharpened to 0.45 m) in combination with LiDAR data was tested for the stress detection with 1089 manually edited reference crowns of kauri and dead and dying trees. Five basic levels of canopy stress symptoms, from non-symptomatic to dead, were further refined for the first symptom stages based on field observations and aerial images. The minimum crown diameter for the use of WorldView-2 attributes for stress detection was defined as 4 m to avoid mixed pixels and to detect dying top branches in smaller crowns. Attributes from only the WorldView-2 image resulted in a correlation of 0.89 (RMSE 0.48, mean absolute error (MAE) 0.34) in a Random Forest regression for crowns larger than 4 m in diameter. This result can be improved to a correlation of 0.92 (RMSE 0.43, MAE 0.31) with additional LIDAR attributes, including intensity values. The selection of attributes confirms the findings from the second study, with an NDVI on near-infrared and red bands as the most important spectral index for the full range of stress symptoms. It also confirms the higher importance of pigment-sensitive indices with green, red and red-edge bands for the detection of first stress symptoms. These initial symptoms are more related to changes in the foliage than the crown architecture.

The results of this thesis present a methodical basis for kauri identification and stress detection using remote sensing data. The methods presented here require further testing and refinement with reference data in other forest areas and should be applied in the full processing chain with automatic crown-segmentation. However, when this has been done, remote sensing methods have considerable potential for automated monitoring of canopy stress symptoms in kauri trees.