Gavish, Y orcid.org/0000-0002-6025-5668, O'Connell, J, Marsh, CJ et al. (4 more authors) (2018) Comparing the performance of flat and hierarchical Habitat/Land-Cover classification models in a NATURA 2000 site. ISPRS Journal of Photogrammetry and Remote Sensing, 136. pp. 1-12. ISSN 0924-2716
Abstract
The increasing need for high quality Habitat/Land-Cover (H/LC) maps has triggered considerable research into novel machine-learning based classification models. In many cases, H/LC classes follow pre‐defined hierarchical classification schemes (e.g., CORINE), in which fine H/LC categories are thematically nested within more general categories. However, none of the existing machine-learning algorithms account for this pre-defined hierarchical structure. Here we introduce a novel Random Forest (RF) based application of hierarchical classification, which fits a separate local classification model in every branching point of the thematic tree, and then integrates all the different local models to a single global prediction. We applied the hierarchal RF approach in a NATURA 2000 site in Italy, using two land-cover (CORINE, FAO-LCCS) and one habitat classification scheme (EUNIS) that differ from one another in the shape of the class hierarchy. For all 3 classification schemes, both the hierarchical model and a flat model alternative provided accurate predictions, with kappa values mostly above 0.9 (despite using only 2.2–3.2% of the study area as training cells). The flat approach slightly outperformed the hierarchical models when the hierarchy was relatively simple, while the hierarchical model worked better under more complex thematic hierarchies. Most misclassifications came from habitat pairs that are thematically distant yet spectrally similar. In 2 out of 3 classification schemes, the additional constraints of the hierarchical model resulted with fewer such serious misclassifications relative to the flat model. The hierarchical model also provided valuable information on variable importance which can shed light into “black-box” based machine learning algorithms like RF. We suggest various ways by which hierarchical classification models can increase the accuracy and interpretability of H/LC classification maps.
Metadata
Item Type: | Article |
---|---|
Authors/Creators: |
|
Copyright, Publisher and Additional Information: | © 2017 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V.. This is an author produced version of a paper published in ISPRS Journal of Photogrammetry and Remote Sensing. Uploaded in accordance with the publisher's self-archiving policy. |
Keywords: | Classification; Machine-learning; Hierarchical models; Random forest; NATURA 2000; Habitat/Land-Cover |
Dates: |
|
Institution: | The University of Leeds |
Academic Units: | The University of Leeds > Faculty of Biological Sciences (Leeds) > School of Biology (Leeds) The University of Leeds > Faculty of Environment (Leeds) > School of Geography (Leeds) > Ecology & Global Change (Leeds) |
Funding Information: | Funder Grant number EU - European Union 308454 EU - European Union 308454 |
Depositing User: | Symplectic Publications |
Date Deposited: | 03 Jan 2018 11:30 |
Last Modified: | 05 Feb 2019 01:38 |
Status: | Published |
Publisher: | Elsevier |
Identification Number: | 10.1016/j.isprsjprs.2017.12.002 |
Open Archives Initiative ID (OAI ID): | oai:eprints.whiterose.ac.uk:125666 |