Interannual spectral consistency and spatial uncertainties in UAV-based detection of boreal and subarctic mire plant communities

Abstract

Abstract

Unoccupied Aerial Vehicle (UAV) imagery is widely used for detailed vegetation modeling and ecosystem monitoring in peatlands. Despite high-resolution data, the spatial complexity and heterogeneity of vegetation, along with temporal fluctuations in spectral reflectance, complicate the assessment of spatial patterns in these ecosystems. We used interannual multispectral UAV data, collected at the same time of the year, from two aapa and two palsa mires in Finland. We applied Random Forest classification to map plant communities and assessed spectral, temporal and spatial consistency, class relationships and area estimates. Further, we used the class membership probabilities from the classification to derive a secondary classification map, representing the second most likely class label per-pixel and an alternative map to account for spatial uncertainty in area estimates. The accuracies of the primary classifications varied between 66 and 85%. The best results were achieved using interannual data, improving accuracy by up to 14%-points when compared to single-year imagery, particularly benefiting classes with lower accuracies. Spectral and temporal inconsistencies in the UAV data collected in different years led to variations in the classifications, notably for the Rubus chamaemorus community in palsa mires, likely due to weather fluctuations and phenology. The transformations from primary to secondary classifications in areas of high uncertainty aligned well with the class relationships in the confusion matrix, supporting the model's reliability. Confidence interval-based adjusted estimates aligned largely with unadjusted area estimates of the alternative map. Our findings support incorporating class membership probabilities and alternative maps to capture spatially explicit uncertainty, especially when spatial variability is high or key plant communities are involved. Our presented approach is particularly beneficial for upscaling ecological processes, such as carbon fluxes, where spatial variability is driven by plant community distribution and where informed decision-making requires detailed spatial assessments.