Classifying kudzu using high-resolution multispectral and Lidar data
Abstract
Mapping distribution of invasive plants is of great importance to enhance monitoring and management activities. Remote sensing (RS) is an increasingly popular tool for mapping vegetation. Kudzu, as an invasive vine, has led to alteration of forest canopy structure, species biodiversity loss, and economic loss. Currently, a method to map the distribution of kudzu using widely available RS data is lacking; this research aims to address this gap. High-resolution multispectral RS imagery, along with Lidar point clouds, digital elevation model, and national land cover data from 2016 (NLCD 2016), were used to detect kudzu. The entirety of Knox County in Tennessee, United States, was used as the study area. Object-based image analysis was used for kudzu classification, and spectral and topographic features, texture,and Lidar-derived metrics were used as predictors. Except for accurately labeled kudzu samples, 10,000 randomly extracted samples were first coarsely labeled based on NLCD 2016. Random forest (RF) was used to predict kudzu objects using these kudzu and coarse-labeled samples. Finally, 300 samples were randomly extracted from the predicted kudzu objects by RF and labeled manually. Four machine learning models were used with the 300 accurately labeled samples to predict kudzu. Among the four models, RF and support vector machine (SVM) were the best two models. The testing Producer's Accuracy for RF and SVM was 89.0% and 76.5%, respectively, while the testing User's Accuracy for RF and SVM was 66.8% and 73.0%, respectively. The errors mainly resulted from misclassification between kudzu and herbaceous plants.Lidar-derived features were the most important predictors, followed by vegetation index and texture features. This research shows the potential in using widely available RS data to detect kudzu. The designed workflow can be easily implemented to map distribution of kudzu or other invasive plants in other regions.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.B11F2393L
- Keywords:
-
- 0410 Biodiversity;
- BIOGEOSCIENCES;
- 0439 Ecosystems;
- structure and dynamics;
- BIOGEOSCIENCES;
- 0480 Remote sensing;
- BIOGEOSCIENCES