Data and resolution requirements in mapping vegetation in spatially heterogeneous landscapes

Abstract

It has been argued that even centimeter-level resolution is needed for mapping vegetation patterns in spatially heterogeneous landscapes such as northern peatlands. However, there are few systematic tests for determining what kind of spatial resolution and data combinations are needed and what the differences in mapping accuracy are when different datasets are omitted or included. We conducted 78 different object-based supervised random forest classifications on a patterned fen and its surroundings in Kaamanen, northern Finland, using remotely sensed optical imagery, topography, and vegetation height datasets from different platforms (unmanned aerial vehicle (UAV), aerial, satellite) with spatial resolution ranging from 5 cm to 3 m. We compared differences in classification performance when we altered (1) classification and segmentation input data and features calculated from the data, or (2) the segmentation scale. We constructed training data with the help of transect-based field sampling and UAV imagery and tested classification accuracy using 412 field-surveyed vegetation plots. The most accurate classifications (75.7% overall accuracy) were obtained when we segmented a 5 cm resolution UAV image with a small segmentation scale and calculated features from all datasets. Classification accuracy was 2.2 percentage points (pp) lower with the most accurate aerial image (50 cm resolution) based classification, and 7.6 pp and 11.9 pp lower with the most accurate WorldView-2 (2 m resolution) and PlanetScope (3 m resolution) satellite image based classifications respectively. Classification accuracies were low (46.7–56.0%) when we used only spectral data from one dataset. The inclusion of gray-level co-occurrence matrix textural features increased classification accuracy by 0.4–12.1 pp and inclusion of multiple datasets by 8.2–25.0 pp. Segmentation scale had a minor effect on classification accuracy (2.5–7.3 pp difference between the finest and coarsest segmentation scale); however, both too small and large segmentation scale might lead to suboptimal classification. The differences in land cover type areal coverage were relatively small between classifications with multiple datasets, but if classifications included features from only one dataset, the differences were larger. We conclude that multiple different optical, topographical, and vegetation height datasets should be used when mapping vegetation in spatially heterogeneous landscapes, and that sub-meter resolution data (e.g. UAV or aerial) are necessary for the most accurate maps. Although UAV data is not essentially needed for classification, it is useful for training dataset construction and especially helpful in areas lacking other sub-meter resolution data.