Tree species identification in ex situ conservation areas using WorldView-2 Satellite Data and Machine Learning Methods: a case study in the Bogor Botanic Garden

Spatially-explicit data on the species composition of forest plants can be an important tool for forest management and conservation. One specific application of these data is for identifying tropical tree species through machine learning techniques to classify satellite remote sensing images. This study aims to examine the ability to use Worldview-2 high-resolution data with various machine learning methods to identify tree species in the Bogor Botanic Garden. Eighteen species from 11 families were selected as samples representing an ecologically and taxonomically diverse data set. Using aggregated image variables, each tree species was found to have different reflectance, texture, and spectral vegetation index variable values. Cluster analysis showed that the 18 tree species could be separated into three clusters that partly reflected taxonomic relationships. Four machine learning algorithms (Support Vector Machine (SVM), Random Forest (RF), K-nearest neighbor (KNN), and Bayesian) were used to predict the species identity of pixels in the image data. A multicollinearity test using a Variance Inflation Factor method reduced the predictor variables from 54 to 9. The highest accuracy (0.96) was observed using SVM, followed by RF (0.91), KNN (0.86), and Bayesian (0.74). The implementation of high-resolution satellite imagery and machine learning for species identification in tropical ex situ plant conservation areas, such as botanic gardens is reported here for the first time.