Detection of Yucca gloriosa in Mediterranean coastal dunes: A comparative analysis of field-based sampling, human interpretation of UAV imagery and deep learning to develop an effective tool for controlling invasive plants

Abstract

Using UAV imagery is a powerful method for monitoring invasive alien plant species (IAPs), particularly when combined with automatic image analysis conducted by artificial intelligence. To this end, we conducted a pilot study on Yucca gloriosa, an invasive species of coastal dunes spread in central Italy. Specifically, we assessed the agreement in quantifying Y. gloriosa cover between field-based sampling and human visual screening of UAV images captured at different altitudes. Additionally, we examined the concordance among different operators both before and after a training procedure, comparing a simpler and quicker approach (referred to as the “envelope” method) against a seemingly more precise but time-consuming method (referred to as the “leaf by leaf” method). In our current study, we discovered a good concordance not only between operators and field sampling but also among operators, particularly when using the “envelope” method. Furthermore, we assessed the performance of deep learning in identifying Y. gloriosa plants in UAV images compared to visual identification by human operators, achieving an overall accuracy of 96 % for images taken at an altitude of 35 m. Our findings suggest that UAV imagery may serve as a valid alternative to field-based sampling for monitoring IAPs, especially when dealing with plants like Y. gloriosa, which have distinctive morphological characteristics that facilitate identification. Consequently, mapping Y. gloriosa on Mediterranean coastal dunes can be effectively accomplished using UAV images, even though an automated machine-based approach, thereby expediting and enhancing the reliability of alien species monitoring and management.