Author(s): Akito Momose; Shuji Iwami; Takayuki Nagaya; Hitoshi Miyamoto

Linked Author(s): Hitoshi MIYAMOTO

Keywords: No Keywords

Abstract: This paper examined machine learning algorithms for classifying riverine landscapes of UAV (Unmanned Aerial Vehicle, i. e., drone) remotely sensed images. The images were taken in a river channel of Kurobe River, central Japan in November 2017. The components of UAV measurements used for machine learning were an RGB, NDVI (Normalized Difference Vegetation Index), and DSM (Digital Surface Model). In addition, an image object, a set of pixels with similar characteristics in RGB images was used as the smallest unit in the land cover classification. The algorithms of machine learning examined in this paper were SVM (Support Vector Machine), RF (Random Forest), and GBDT (Gradient Boosting Decision Tree). True classes of trees, grasses, bare gravel/sand beds, and water surface in a part of the river channel were given as a training dataset. The results in accuracy evaluation using the F-measure showed that DSM is effective for improving the identification of trees, and NDVI for vegetation and water surface. It could be concluded that, for the test section examined in this paper, RF was the most effective machine learning algorithm in terms of the detection accuracy and CPU time efficiency.

DOI:

Year: 2018