Sentinel-2 MSIを用いた竹林の分類—四つの分類アルゴリズムにおける観測時期と観測波長帯の評価—

抄録

The expansion of abandoned bamboo forests is a significant issue in Japan. Remote sensing is an effective technique for mapping land cover over large areas. In order to accurately estimate the distribution of bamboo forests using the MultiSpectral Instrument (MSI) on board the Sentinel-2 satellite, four classification algorithms (random forest, support vector machine, multilayer perceptron, and gradient boosting decision tree) were evaluated, with a particular focus on the seasonality of classification accuracy and the importance of the spectral band. A land cover map of Kochi City was generated for each month using MSI data and reference maps. The performance of the algorithms in the bamboo forests was assessed based on the seasonal profiles of precision, recall, and F-measure. Permutation feature importance was used to measure the importance of spectral bands. The classification accuracy of the random forest classifier was found to be slightly superior to that of the other algorithms, although the difference was smaller in the case of the gradient boosting decision tree. Classification accuracy was high in May and remained relatively high from May through August. All algorithms considered bands 11 and 12 in the short-wavelength infrared region to be significant. In addition, bands 6 and 7 in the red-edge spectral region were of nearly equal importance in the support vector machine, multilayer perceptron, and gradient boosting decision tree algorithms. The study also created seasonal profiles of misclassification of bamboo forests as other land cover types, such as evergreen broadleaf forests or deciduous broadleaf forests. These findings could be used to improve the accuracy of bamboo forest mapping using optical remote sensors.