Interpretable segmentation and naturalness classification of forests using the Canopy Height Model

Abstract

As increasing amounts of ecological data from remote sensing is available, so does the opportunities to utilize this data also increase. One such opportunity is to classify the naturalness of forests automatically from data instead of using costly field-surveys, in the purpose of more efficiently managing natural ecosystems. Doing this using AI-based methods in an interpretable way is the topic of this thesis. Interpretability is of utmost importance in order to improve testing and verifying of models as well as to provide transparency and understanding to affected parties, such as forestry companies. Using the Canopy Height Model (CHM), i.e. laser-scanned height data, as single source of data, three segmentation algorithms are evaluated on the task of creating segments of forest from CHM-rasters. A lack of ground-truth data in the evaluation is overcome by using human feedback as preference-based pairwise comparisons to evaluate and optimize. A framework to collect such data has been developed in the shape of a web application to assure convenience and efficiency in collection of data. Based on this data, one superior algorithm is selected and further optimized, to then be applied in tandem with a previously developed naturalness classifier in order to map the naturalness of forests. The result is a CHM–into–naturalness pipeline with a 75% naturalness classification accuracy, evaluated on a 50 × 50 km area in Southern Sweden, only performing segmentation and classification on 640 × 640 m tiles of the CHM at a time. This method is shown to have both big potential but also some vulnerabilities, among them being the tile sizes, and remedies are suggested and discussed. A certain mismatch in how the model was trained and how it is now evaluated introduces some uncertainty to the evaluation, and possible improvements are discussed. In summary, the thesis lays the ground-work for the topic and suggests important aspects to iterate upon, giving a direction for future research.