The aim of this research was to produce a time series analysis of change in native woody vegetation in selected sheep and beef farming regions using satellite imagery attained from the Landsat programme. This was proposed to be achieved by training a machine learning classifier to accurately estimate vegetation attributes in modern imagery and applying the produced models to historic imagery that predates the New Zealand Land Cover Database. Training machine learning classifiers to recognise the desired suite of land cover classes in modern satellite imagery proved to be a challenging task in itself, and due to image incomparability issues, applying these trained classifiers to historic satellite imagery and quantifying accuracy was not possible in the given time frame. As the challenges in achieving the original aims of this research project were realised, lines of enquiry were altered to more thoroughly investigate novel areas of the workflow where no sufficiently detailed literature exists.

This thesis describes the data preparation and classification methods developed as a foundation on which research into data comparability solutions and classification optimisation methods can be built. Concisely, the three most important outcomes of this work were:

1. Development of a method for preparing classifier training datasets compatible with both the source data (Landsat imagery) and the classifier (a Convolutional Neural Network based hybrid method implemented through Trimble eCognition) for achieving optimal classification accuracy.
2. Development of a classification framework that draws a compromise between the two aims of classifying land cover with high accuracy and classifying land cover with environmentally relevant detail.
3. Development of a functional machine learning classifier able to detect the desired land cover classes in Landsat imagery, including those that are not visible to the human observer.

Importantly, this thesis also describes the major barriers to further development of a method for producing an accurate time series of land cover change. The two most important problems encountered are:

1. Classifier models trained to detect land cover classes with modern Landsat imagery did not achieve any level of useful classification accuracy when applied to historic image datasets.
2. Accuracy assessment of classified maps produced by application of pre-trained classifier models to significantly older datasets proved to be incalculable through conventional methods.

In overview, this thesis should serve as an easily digestible resource to assist in future development of a software solution that can produce classified, time-series maps of native woody vegetation on New Zealand’s sheep and beef farmland at a low cost. The intended purpose of this software is to assist Beef and Lamb New Zealand in achieving the goals set out in their Environment Strategy and Implementation Plan so that they can better support their farmers in acting as effective kaitiaki of the land and remaining compliant and self- regulating as environmental policy in the agriculture sector develops.