Exploring the application of deep learning techniques to facilitate the management of invasive species

Biological invasion threatens biodiversity protection and ecosystem services and is considered one of the major threats to the long-term success of mineland restoration projects. Especially due to the unrestrained growth and significant dispersal capacity of invasive species that jeopardize not only the recovery areas but also pose risks to neighboring environments. Despite the urgency, combating invasive species is still conducted in a manner incompatible with the need for effective large-scale monitoring. A significant challenge has been developing methods capable of detecting these species during the early stages of invasion and monitoring the population on a large scale.

The utilization of refined data, particularly those collected with multi and hyperspectral sensors, has been a focal point for species detection in ecological research, particularly in environments with higher diversity. However, the adoption of such approaches is not yet widespread among environmental monitoring companies, primarily due to challenges associated with costs and the complexities of data collection. Using deep-learning algorithms can than facilitate species detection with simple RGB images providing more possibilities to ecological studies in this field, while improves application as simplifier the process of data acquisition to ecosystems managers. For this purpose, this work used a deep-learning model using RGB images to detect two invasive species Melinis minutiflora Beauv. (Poaceae) and Muntingia calabura L. (Muntingiaceae) in mining restoration sites in the eastern Amazon. Unoccupied aerial systems image data of a waste pile was collected with a total size of approximately 108 ha.

The applied methodology was able to differ invasive species in our study site and the spatial distribution map generated revealed hotspots of M. minutiflora and M. calabura in the restoration area. The detection of these species using RGB images underscores the potential of deep learning to map invasive species and provides a more accessible way for monitoring on a larger scale. In conclusion, our results contribute to improve efficiency of large-scale monitoring of invasive species in restoration projects. By facilitating data collection and highlighting the potential for economically viable management, our findings provide a valuable perspective for stakeholders engaged in enhancing invasive species management practices.