Assessing the Efficacy of Phenological Spectral Differences to Detect Invasive Alien <i>Acacia dealbata</i> Using Sentinel-2 Data in Southern Europe

Abstract

Invasive alien plants are transforming the landscapes, threatening the most vulnerable elements of local biodiversity across the globe. The monitoring of invasive species is paramount for minimizing the impact on biodiversity. In this study, we aim to discriminate and identify the spatial extent of Acacia dealbata Link from other species using RGB-NIR Sentinel-2 data based on phenological spectral peak differences. Time series were processed using the Earth Engine platform and random forest importance was used to select the most suitable Sentinel-2 derived metrics. Thereafter, a random forest machine learning algorithm was trained to discriminate between A. dealbata and native species. A flowering period was detected in March and metrics based on the spectral difference between blooming and the pre flowering (January) or post flowering (May) months were highly suitable for A. dealbata discrimination. The best-fitted classification model shows an overall accuracy of 94%, including six Sentinel-2 derived metrics. We find that 55% of A. dealbata presences were widely widespread in patches replacing Pinus pinaster Ait. stands. This invasive alien species also creates continuous monospecific stands representing 33% of the presences. This approach demonstrates its value for detecting and mapping A. dealbata based on RGB-NIR bands and phenological peak differences between blooming and pre or post flowering months providing suitable information for an early detection of invasive species to improve sustainable forest management.

Keywords

• invasive alien plants
• remote sensing
• phenology
• machine learning