Combining multispectral and texture imagery features to assess health condition in priority riparian forests by means of unmanned aerial systems

Expansion of damaging pests and pathogens is a reality which, together with rapid global change, is arguably the greatest contemporaneous challenge to sustainable forestry and the continuing function of forest ecosystems. Alnus glutinosa (black alder) woodlands are priority riparian forests for conservation at European Scale (Habitat 91E0* of Habitat Directive 43/92/CEE), due to their key ecological functions (such as N fixation, wildlife habitat) and ecosystem services provision (e.g. improvement of water quality). Recently, substantial declines in alder stands have been observed along streams in Europe. A major driver has been the invasive oomycete pathogen Phytophthora alni species complex, with damages widespread across Europe and even in some parts of North America. This is critical, not only due to disproportionate ecological importance of riparian forests in relation to their surface area extent but also due potential impacts to other forest species. Proper management requires accurate assessment of forest status and novel remote sensing devices offer increasing opportunities to overcome high labour costs and time-consuming travels, typical of field based monitoring. The mapping of the defoliation caused by the disease is particularly challenging in high density ecosystems with high spectral variability due to canopy heterogeneity. The use of Unmanned Aerial Vehicle (UAV) data for such tasks might be particularly advisable due to its high resolution, acquisition flexibility and cost efficiency in the field. In this study, Alnus glutinosa decline was assessed by classifying four different health condition levels (healthy, dead, and defoliation under and below a 50% threshold), previously attributed through individual tree field sampling. A combination of multispectral Parrot Sequoia and RGB-UAV-data were analysed using Random Forest (RF) and a simple and robust three-step logistic modelling approaches to identify the most relevant predictors and keep the models parsimonious. A total of 34 remote sensing (RS) variables were included in the study, including a set of vegetation indices (VI), texture features from NDVI and DSM (Digital Surface Model), topographic and DAP (Digital Aerial Photogrammetry)-derived structural from Digital Surface Model (DSM) at crown level. The four level health condition classification achieved an overall classification accuracy of 67%. On the other hand, the confusion matrix computed from the three logistic models using leave-out cross-validation method achieved an overall accuracy of 76% when using four level health condition classification. Our results offer an alternative robust classification method to forest and conservation managers for the rapid and effective assessment of areas affected by the disease in their planning of control and restoration measures aimed at reducing these forests vulnerability and black alder mortality with potential application to other species.