Detecting spatial patterns of land cover and methane fluxes with remote sensing in Pallastunturi, Finland

• Main Author: Rauth, Ella
• Other Authors: Helsingin yliopisto, Matemaattis-luonnontieteellinen tiedekunta, University of Helsinki, Faculty of Science, Helsingfors universitet, Matematisk-naturvetenskapliga fakulteten
• Format: Master Thesis
• Language: English
• Published: Helsingin yliopisto 2022
• Subjects: random forest; remote sensing; land cover; methane; peatland; Datatieteen maisteriohjelma; Master's Programme in Data Science; Magisterprogrammet i data science; ei opintosuuntaa; no specialization; ingen studieinriktning; Pallastunturi; Tundra
• Online Access: http://hdl.handle.net/10138/344123

Northern peatlands are a large source of methane (CH4) to the atmosphere and can vary strongly depending on local environmental conditions. However, few studies have mapped fine-grained CH4 fluxes at the landscape-level. The aim of this study was to predict land cover and CH4 flux patterns in Pallastunturi, Finland, in a study area dominated by forests, peatlands, fells, and lakes. I used random forest models to map land cover types and CH4 fluxes with multi-source remote sensing data and upscaled CH4 fluxes based on land cover maps. The random forest classifier reliably detected the same land cover patterns as the CORINE Land Cover maps. The main differences between the land cover maps were forest type classification, misclassification between neighboring peatland types, and detection of sparsely vegetated areas on fells. The upscaled CH4 fluxes of sinks were very robust to changes in land cover classification, but shrub tundra and peatland CH4 fluxes were sensitive to the level of detail in the land cover classification. The random forest regression performed well (NRMSE 6.6%, R2 82%) and predicted similar CH4 flux patterns as the upscaled CH4 flux maps, despite predicting larger areas that act as CH4 sources than the upscaled CH4 flux maps. The random forest regressor also better predicted CH4 fluxes in peatlands due to added information about soil moisture content from the remote sensing data. Random forests are a good model choice to detect landscape patterns and predict CH4 patterns in northern peatlands based on remote sensing and topographic data.