RGB vs. Multispectral imagery: Mapping aapa mire plant communities with UAVs

Abstract

Plant communities of mires can be linked to important ecological processes, such as carbon storage and gas fluxes. As indicators of ecosystem dynamics, knowledge about their distribution and condition can support ecosystem assessment. Mapping mire vegetation enables monitoring at ecosystem-scale, which can be done with UAVs (Unmanned Aerial Vehicles). Depending on the mounted sensor and the spectral signals recorded, various attributes of plant communities can be retrieved. However, it is uncertain to what extent plant communities can be derived, as mapping vegetation on detailed level remains challenging due to overlapping spectral signatures of plant species. Advancing technology offers the choice between low cost RGB and multispectral sensors as well as a variety of classification methods to overcome these challenges. Therefore, we used K-means unsupervised classification and Random Forest supervised classification with different input variables to map microtopographical patterns and plant communities of two aapa mires as resolved by hierarchical clustering. This extensive approach allowed the assessment of both classifier’s strength and weaknesses, as well as the criteria of selecting suitable input data. UAV- RGB and multispectral imagery with associated spectral and topographical indices of both 0.05 m and 0.30 m spatial resolution were used for the K-means method. We assessed the relationship between these generated spectral classes and plant community clusters. The clusters further served as training and validation labels for to classify the high resolution, multispectral UAV-data (0.05 m) using Random Forest. Our study demonstrates that maps reflecting microtopograpical patterns and a wetness gradient can be produced with low-cost RGB imagery and unsupervised classification. Despite this linkage to plant communities, enhanced and detailed maps of plant community distribution can only be achieved with multispectral data and robust machine learning techniques. Random forest classifications showed good overall accuracies (0.59 – 0.82) in mapping microtopographical patterns and plant communities based on hierarchical clustering of vegetation data. While the strength of both classifiers lies in the distinction of bog hummock communities, classification performance was weaker between different transition and lawn community types. Casual misclassifications occurred also for communities along the transition of microtopographical patterns. The main obstacle for accurate mapping remains the overlap of spectral signatures from species and spectral noise originating from wetness in mires that lead to misclassification. Future studies addressing plant community and diversity mapping should therefore consider the origin of spectral variation with further sensors.