Machine learning-based classification of woodland bitter vine (Mikania micrantha Kunth)

Abstract

The woodland invasive species bitter vine (Mikania micrantha Kunth) is widely distributed in southern Asia. Since it adversely affects vegetation growth and ecological functions, it is imperative to develop an automated detection method to investigate the species’ expansion in large areas. We obtained training and test samples in the Zengcheng woodland, Guangzhou Province, China, and used aerial imagery to classify bitter vine using machine learning algorithms. The results suggested that visual interpretation was suitable for bitter vine classification but had slightly lower accuracy than methods using field-based training samples. The random forest (RF) algorithm proved most efficient for large-scale detection with the highest overall classification accuracy (99.62% for the test set and 99.92% for the training set) and Kappa coefficient (0.9878 for the test set and 0.9975 for the training set). Although the backpropagation neural network-based deep learning (DL) approach showed uncertainties and errors in this study, a more advanced DL method could be suitable for this task but requires further research. Bitter vine occurs predominantly in disturbed areas with low vegetation cover. Therefore, monitoring, silvicultural measures, and manual removal are necessary. Since bitter vine is highly distinguishable in images, future studies should further investigate supervised classification methods to map and monitor this species.