Search

Abstract Hyperspectral data and machine learning offer great potential for identifying valuable open ecosystems. Due to the large volume of data, preprocessing of hyperspectral images must involve dimensionality reduction. The main goal of this study was to test the effectiveness of various types of feature reduction (feature selection and feature extraction) when performing classification using the Random Forest algorithm. A comparison was conducted between two ecosystems - heathlands and mires protected as Natura 2000 habitats. Two transformations of feature extraction were chosen, namely Minimum Noise Fraction (MNF) and Principal Component Analysis (PCA), while Linear Discriminant Analysis (LDA) was used as a feature selection method. It was proven that irrespective of the class type, accuracy is higher with the feature extraction method (mean F1 accuracy of 0.922) than with feature selection (mean F1 accuracy of 0.787). At the same time, no significant differences in accuracies were found between the MNF and PCA methods. Although LDA resulted in lower accuracies (0.816 for heathland and 0.750 for mires), the method could also be used due to relatively high F1 values. The effectiveness of the LDA method for feature reduction in open ecosystem identification was confirmed for the first time for open natural vegetation.

Abstract Aerial hyperspectral and multispectral satellite data are the two most commonly used datasets to identify natural and semi-natural vegetation. However, there is no documented analysis based on data from several areas concerning the difference in the classification accuracy of non-forest Natura 2000 habitat with the use of aerial hyperspectral and satellite multispectral data. Also, there is no recommendation, on which habitat can be classified with sufficient accuracy using free multispectral images. This study aimed to analyse the difference in classification accuracy of Natura 2000 habitats representing: meadows, grasslands, heaths and mires between data with different spectral resolutions and the results utility for nature conservation compared to conventional maps. The analysis was conducted in five study areas in Poland. The classification was performed on multispectral Sentinel-2 (S2) and hyperspectral HySpex (HS) images using the Random Forest algorithm. Based on the results, it can be stated that the use of HS data resulted in higher classification accuracy, on average 0.14, than using S2 images, regardless of the area of the habitat. However, the difference in accuracy was not constant, varying by area and habitat characterisation. Greater differences in accuracy were observed for areas where habitats were characterised by high α-diversity or β-diversity. The HS and S2 data make it possible to create maps that provide a great deal of new knowledge about the distribution of Natura 2000 habitats, which is necessary for the management of protected areas. The obtained results indicate that by using S2 images it is possible to identify, at a satisfactory level, alluvial meadows and grassland. For heaths and mires, using HS data improved the results, but it is also possible to acquire general distribution of these classes, whereas HS images are obligatory for mapping salt, Molinia and lowland hay meadows.