2021 Volume 23 Issue 2 Pages 261-278
In this study, we developed a vegetation discrimination method based on satellite remote sensing, topographical information, and machine learning for creating vegetation maps along the Arakawa River. First, satellite images were processed by object-based classification. Then spectral information from the satellite imagery, the vegetation index, and topographical information were added to each object. Vegetation discrimination models were created by machine learning algorithm (Random Forests, Support Vector Machine), and individual objects were adapted to the vegetation types. Comparison of the results with existing vegetation maps confirmed that the use of topographical information improved the accuracy of discrimination. When the amount of training data for machine learning was reduced to 10%, the classification accuracy of the Support Vector Machine algorithm was higher than that of the Random Forests algorithm. In addition, the accuracy of vegetation discrimination was reduced when vegetation types with a small number of data were generated in the training data. When using field survey data to classify vegetation with a small sampling rate, it was shown that the number of survey sites per vegetation type needs to be controlled so that the number of training data does not become too small.
