Back to search

University of British Columbia

Classifying wetlands using random forest machine learning, airborne light detection and ranging and Earth observation satellite data in the Okanagan basin, British Columbia

Abstract

dc:description

Despite wetlands being critical components of healthy functioning landscapes and providing valuable ecosystem services, they are being lost at alarming rates due to human influence. Their dynamic nature and varying hydrology and vegetation make identification and classification challenging. Recent advancements in satellite remote sensing and light detection and ranging (LiDAR) provide opportunities to predict and map wetland classes at regional scales. In the semi-arid region of the Okanagan Basin, wetlands are rare biodiversity hotspots that provide critical habitat for many species at risk. Building on existing wetland inventories that have limited coverage, a random forest probabilistic model was developed to predict, classify and map wetlands in the Okanagan at a 10 m spatial scale. In total, 22 covariates representing multispectral and synthetic aperture radar metrics derived from Sentinel-2 and Sentinel-1; topography and vegetation derived from LiDAR; and ancillary geospatial data were used to classify wetlands. The model was trained using an existing wetland database and provincial datasets to predict the probability of each pixel belonging to the following six-classes: fen, marsh, shallow-water, swamp, upland, and open-water. Model performance was evaluated using a confusion matrix and had an overall accuracy of 84.8%. The model predicted that 313.9 km² (3.6%) of the 8,635 km² study area represented areas where wetland probability was ≥ 50%. Marshes were the most commonly occurring wetland (159.0 km²) followed by swamp (150.3 km²), shallow-water (3.7 km²), and fen (0.9 km²). The most important predictor variables for wetlands were slope, distance from streams, probability of depression, number of days above five degrees Celsius, topographic position index, seasonal change in the normalized difference vegetation index, standard deviation of vegetation height, and the red band from Sentinel-2. The wetland model developed here identified and classified new wetlands and provided a comprehensive inventory of wetlands in the Okanagan using a replicable approach with publicly available data. The resulting wetland inventory can help inform regional wetland conservation and management and will serve as an important baseline for land use planning and climate change mitigation.

Degree

thesis:*
Name thesis:degree_name
Master of Science - MSc
Level thesis:degree_level
master's
Discipline thesis:degree_discipline
Earth and Environmental Sciences
Grantor dc:publisher
University of British Columbia
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Deenik, Kristina
Contributors dc:contributor
  • Bourbonnais, Mathieu

Rights

dc:rights
Statement dc:rights
  • Attribution-NonCommercial-NoDerivatives 4.0 International
Language dc:language
eng

Identifiers

dc:identifier.*
Handle dc:identifier
http://hdl.handle.net/2429/81703

Chain of custody

source
Harvested from
University of British Columbia
Base URL
circle.library.ubc.ca/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Deenik, Kristina. Classifying wetlands using random forest machine learning, airborne light detection and ranging and Earth observation satellite data in the Okanagan basin, British Columbia. master's thesis, University of British Columbia, 2022. http://hdl.handle.net/2429/81703