Modelling aapa mire flark coverage with multi-resolution remote sensing data

Abstract

Peatlands have degraded worldwide significantly because of human influence and global warming. The degradation and drying of peatlands have increased the need to monitor the state of peatlands, in which remote sensing (RS) methods have proven to be efficient. In the boreal aapa mire zone, the peatlands are characterized by patterned wet flark-string areas in the centres of the peatlands. The water-filled flark areas indicate the ecological state of aapa mires.

In previous studies, the peatland wetness has been modelled in numerous ways, though fewer studies have focused on the flark area. In this study, the flark coverage of aapa mires was predicted with satellite data across five sites. The more specific aim was to test how different spatial resolution sensors (Landsat 8-9; 30 m, Sentinel-2; 10 m, PlanetScope; 3 m) and spectral datasets affect the accuracy of flark coverage prediction. The spectral datasets used were with and without shortwave infrared (SWIR) data. Additionally, the effect of seasonal and site-specific differences on model performance was tested. Lastly, the flark coverage was upscaled from unmanned aerial vehicle (UAV) imaged areas to the whole mire areas. The high-resolution UAV images were used as ground truth data to classify the flark coverages, which were used as training data for regression models. The flark coverage was predicted by random forest regression, in which single bands and indices from satellite data were used as predictor variables.

All sensors provided accurate flark coverage prediction results, with some differences between the sensors: Landsat 8-9 (pseudo-R² 32−84%, root-mean-squared error (RMSE) 10−18%), Sentinel-2 (R² 61−92%, RMSE 6−14%), and PlanetScope (R² 46−92%, RMSE 6−17%). Using the SWIR data did not improve model accuracies. The differences between the seasons were modest, yet the models of early summer were generally more accurate than late summer models. The site-specific differences were notable, as the three best modelling sites had similar, distinctive flark-string patterns. Moreover, the single-site models were highly accurate, whereas the multi-site models had more variance and less accurate results, especially in the models of late summer data. Lastly, the upscaling of the flark coverage to whole mire areas was successful according to visual interpretation. This study demonstrates that flark coverage can be successfully predicted with multi-resolution satellite data paired with UAV data. Furthermore, these methods can be utilized in the monitoring of peatlands' state.