Abstract
In this study, we aimed to assess bamboo distribution via pixel-based image classification of aerial photographs. The study area was the Tachibana Town, Yame City, Fukuoka Prefecture. Aerial photographs taken in June, 2016 and October, 2021 were used for the analysis. A model was constructed to classify the images into “bamboo,” “forest,” and “others” categories using the random forest machine learning method. The overall accuracy of the model was 96% in June, 2016 and 91% in October, 2021. Feature importance analysis revealed visible blue light and red light bands from ortho-rectified aerial photographs processed with a 51×51 averaging filter as the most important features to determine bamboo distribution. The high accuracy observed in June, 2016 was possibly due to the photographs being taken during the bamboo leaf-shedding period. However, in October, 2021, many photographs in the “bamboo” and “forest” categories were misclassified. Overall, these findings suggest that aerial photographs taken during the leaf-shedding period are the most suitable to assess bamboo distribution. As the Fukuoka Prefecture conducts aerial photography of the same area every five years, long-term changes in bamboo distribution can be effectively monitored using aerial photographs.
https://ror.org/05qg1fd04 
