A SVM-based model for predicting distribution of the invasive tree Miconia calvescens in tropical rainforests

Advances in spatial and spectral resolution of sensors can be useless to detect directly understory and subcanopy species but make remote sensing of some ecosystems components increasingly feasible. We propose to use support vector machines (SVM) to integrate multisource-derived biophysical descriptors (overstory plant species, physiography and climate) for the indirect detection of the small invasive tree Miconia calvescens in tropical rainforests on the island of Tahiti (South Pacific). Our model consists in classifying overstory plant species from an optical Quickbird scene, with the output then used in a subsequent fusion process with digital elevation model (DEM) extracted physiographic and climatic descriptors. A range of accuracy metrics was calculated to assess the SVM-based model which widely outperforms the commonly used GARP model. Biophysical descriptors alone are necessary and sufficient to explain M. calvescens distribution and suggest that the potential invaded area is currently saturated in our study site on Tahiti. Rainfall, elevation and slope steepness are the major variables explaining the species distribution. In addition, our results show that morning insolation plays a critical role on M. calvescens height whether it is restricted to the subcanopy or reaches the forest surface. The model can be used to map the potential distribution of M. calvescens in areas where it has been recently introduced and rapidly spreading, such as in the Hawaiian islands, New Caledonia and Australia or in other French Polynesian islands. It also may be adapted to detect other species (plants or animals, alien invasives or rare endemics) in the understory and subcanopy of forest ecosystems.

► We propose to use support vector machines (SVM) for species distribution modelling.
► SVM is compared to the widely used GARP model to map the invasive tree Miconia calvescens on Tahiti.
► Our analysis shows that SVM is more accurate than GARP.
► This research can contribute to improve ecological sustainability of vulnerable landscapes.