A deeper look: The development of global peat depth datasets and subsequent carbon stock estimates
Date
2025
Authors
Skye, Jade Erin
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Peatlands are important carbon stores which are being destabilised by anthropogenic activity and are sensitive to climate change. To faithfully assess the carbon stored in peatlands and to model their responses to future climate scenarios, it is essential to have accurate information on peat depth. Presently, however, observations of peat depth are insufficient for conducting these tasks at the global scale. Thus, the goal of my thesis is to accurately generate a global distribution of peatland depth and use that distribution to estimate how much carbon is stored within them. The first step was to create Peat-DBase, the largest database of harmonised peat depth measurements at the global scale. Peat-DBase was then used as the basis of training and testing data for PeatDepth-ML, a machine learning-based modelling framework designed to predict peat depths globally. I created PeatDepth-ML by adapting an existing modelling framework that was designed to predict peatland spatial extents by including new datasets of environmental variables that may drive or indicate peat formation, updating the cross-validation procedures used for model testing, and adding a custom scoring metric to the model to assist in predicting deeper peat depths. I then used PeatDepth-ML to produce a spatially continuous global map of peatland depths. Inspection of Peat-DBase revealed regional data gaps, such as in the Tropics, and potential sampling biases in peat depth measurements, e.g. the collecting of a single peat core to represent the depth of an entire peatland wherein depth could be varying significantly or the presence of multiple peat cores with highly varying depths over small spatial scales. The impact of Peat-DBases's regional biases on PeatDepth-ML's predictions was assessed by calculating a metric describing the predictions area of applicability. To test the sensitivity of PeatDepth-ML to some aspects of sampling bias, a bootstrapping method was developed to create multiple training datasets from Peat-DBase. Running PeatDepth-ML on the bootstrapped datasets showed that model behaviour could vary significantly in response to changes in the training data, particularly at the regional scale. When compared to other estimates in the literature, PeatDepth-ML achieved a similar or improved level of performance and is of better overall quality because of its global reach and continuous representation of peat and non-peat regions without the use of an independent peatland extent map. However, PeatDepth-ML demonstrated a tendency to predict towards the mean peat depth of its training data, which was relatively shallow possibly due to the inclusion of non-peat data, which was included to allow the model to predict over all regions. Performing simple carbon stock calculations using PeatDepth-ML’s results produced estimates that are in line with those previously published. Collectively, Peat-DBase and PeatDepth-ML are cohesive global datasets of peat depth that can aid future peatland research and policy endeavors.
Description
Keywords
Peat depth, Machine learning