Spectral discrimination of prosopis glandulosa (mesquite) in arid environment of South Africa: Testing the utility of in situ hyperspectral data and guided regularized Random forest algorithm

Prosopis is an evergreen invasive alien species that thrives in the world's most arid and semi-arid environments. Although initially introduced for its benefits such as sand dune stabilization, furniture production and fodder for livestock, over time mesquite showed to have negative impacts socially, economically and ecologically. Hence, in 2004 Prosopis was rated the world's top 100 least wanted species by the IUCN. In South Africa one of the most invaded areas is the Northern Cape Province where it grows amongst other acacia species. Methods involving chemical, mechanical and biological control have been tried and tested with little success due to lack of knowledge on key aspects of the invasion dynamics. Therefore, up to date temporal and spatial information about mesquite invasion is crucial for creating sustainable management plans. This study aimed to test the use of hyperspectral remote sensing to spectrally discriminate Prosopis glandulosa from other coexistent acacia species in the area. In situ hyperspectral data was collected in March 2015 using a Spectral Evolution spectroradiometer. In addition, a new developed guided regularized random forest (GRRF) algorithm was used for variable selection in identifying key wavelengths that accurately discriminate among the tree species. The results show that the new GRRF algorithm was able to reduce the problem of high dimensionality associated with hyperspectral data by selecting key wavelengths within the visible and near infrared regions. The selected wavelengths (n = 11) were then used as input variables in the random forest classifier to discriminate between the four species which yielded an overall accuracy of 88.59% and kappa value of 0.85. Overall, this study has revealed that it is possible to map Prosopis glandulosa from other coexistent acacia species and it is worth considering the new developed GRRF ensemble as a robust method for hyperspectral variable selection and high dimensionality reduction.