Remote Sensing and Object-Based Image Analysis of Restored Peatlands

Sammanfattning

Peatlands globally act as carbon sinks and suffer from a long history of exploitation, leaving them drained and vulnerable to peat decomposition. Restoring peatlands is gaining popularity and funding, but evaluating the success of a restoration is often considered to be a lengthy and expensive process. Restoration success can be gauged using vegetation, especially through plant functional types as these change in composition according to their environment, such as water table height. Remote sensing using Unmanned Aerial Vehicles (UAVs) and plant functional types provide a promising alternative for peatland monitoring. My goal in this study was to evaluate how time since restoration and number of plant functional types affect classification accuracy. I also studied the effect of cross-site model application on classification accuracy. Seven peatlands across south-centre Sweden, ranging from non-restored to 25 years since restoration, were mapped using UAVs. Vegetation datasets consisting of plant functional types were constructed through simple random sampling in field surveys and office interpretation. Each dataset provided three categories of increasing class resolution, as the number of possible classes increased for each category. Classification accuracies were relatively high considering few modifications were made to the default algorithm parameters. Overall accuracies averaged 95 % (Range 88.7 - 99.5 %), 76 % (Range 66.7 - 79.9 %), and 71 % (Range 58.8 - 81.5 %) for the three categories of increasing class resolution. The cross-site model application test revealed that the cross-site model application always performed worse than the same-site model, -7.6 % overall accuracy averaged across all sites and categories. The number of classification classes appeared to be the largest predictor of classification accuracy. The number of classes could also be assumed to be affected by restoration age, which did not show any clear relationship to overall accuracy. The study concluded that applying the same UAV and classification peatland mapping method could provide similar classification accuracies across a large range of restored sites, if the effect of different numbers of classification classes is addressed. Finally, the cross-site model application concluded that same-site classification models are to be preferred for higher classification accuracy.