A Machine Learning model for lianas classification based on multi-source UAV-based Remote Sensing datasets
Abstract
As a major sink in the global carbon cycle, tropical forests store nearly 30% of global terrestrial carbon through photosynthesis, contributing to 40% of the global terrestrial carbon sink. However, lianas (woody vines) impact the carbon balance of tropical forests by affecting the host tree's growth and survival. Selective liana-cutting is a highly cost-effective natural climate solution and a profitable forestry intervention that should be standard practice in managed forests where liana infestations are prevalent. Therefore, accurate and efficient classification and mapping of liana infestations contribute to forest regulation and decision-making regarding liana-cutting. However, to know how severe liana infestation is, one usually relies on manpower either in the field or through visual interpretation based on aerial photos, which are both time-consuming and high-cost. This project develops a Random Forest classification model to predict or classify liana infestations in tropical dry forests based on multi-source UAV-based remote sensing datasets. To train the model, around 130 pixels are manually labeled for each class (i.e., lianas, healthy canopies, and forest gaps) to extract metrics from remote sensing datasets. In terms of the validation for the model's performance, the confusion matrix shows an accuracy of around 86%, and the model is also evaluated by the comparison between prediction result and careful visual interpretation for the unknown area. The results show that the developed model has the potential to classify lianas accurately, which helps support selective liana-cutting decision-making and forest management.