Value of hyperspectral data for wall to wall wetland vegetation mapping in heterogeneous landscapes

Abstract Remote sensing is increasingly employed today for the creation of wall-to-wall vegetation maps of protected areas. The main objective of this study was to evaluate whether it is feasible to develop a remote sensing-based wall-to-wall vegetation map of wetlands that is useful for nature conservation purposes without hyperspectral data. The research was conducted in the Narew National Park in north-eastern Poland. In 2020, three types of airborne data covering an area of 325 km2 were acquired: hyperspectral, Light Detection and Ranging data (LiDAR), and from visible and near infrared range (RGBN) imagery. At the same time, botanists collected 1,921 reference polygons for training and validating the machine learning model. Then, classifications were conducted using the CatBoost algorithm with four different scenarios created from the mentioned datasets. The results indicate that, based on hyperspectral and LiDAR data, it is possible to produce a wall-to-wall vegetation map with an overall accuracy of 0.82. When hyperspectral data are replaced by less expensive and widely accessible RGBN data, accuracy declines by 0.10; this reduction decreases to 0.09 if textural features derived from the RGBN data are additionally incorporated into the classification. In scenarios where only LiDAR data are available, vegetation classification accuracy is much lower, reaching 0.63. Comparison of the remote sensing maps with a vegetation map created through field mapping methods indicates that, irrespective of whether hyperspectral or only RGBN and LiDAR data are available, remote sensing maps contribute substantial new knowledge regarding wetland vegetation in spatially heterogeneous landscapes and are highly useful for their conservation management.