Examination of classification methods and feature selection on forest type mapping by object-based image analysis

現在、リモートセンシングデータを使用した迅速な相観植生図の作成が期待されている。本論ではSPOT5データを使用し、佐渡地域における森林タイプ分類を試みた。画像分類にはオブジェクトベース画像分類を用い、NN（Nearest Neighbor）法とCART（Classification and Regression Trees）法の２種類の分類手法や、初期段階で分類に組込む特徴量を変えた分類結果を比較した。本研究では、初期段階で組込む特徴量の少ないNN 法が最も良い分類精度を示した。NN 法での、初期段階で組込む特徴量の数の違いによる分類結果を比較すると、特徴量を多く組込んだ方が分類精度は低くなっており、特徴量の多さが分類精度の向上に必ずしも結び付くわけではないということが分かった。一方、CART 法では、初期段階で組込む特徴量の数を変化させても、選択される特徴量が限定されており、両者に差は認められなかった。これらの結果から、分類手法に応じて、初期段階で分類に組込む特徴量の各クラスの分離に対する有効性を精査し、事前に選別した方がいい場合があることが確認された。Remotely sensed data would be expected to be beneficial for vegetation mapping. In this paper, we reported forest type mapping using SPOT5 data in Sado Island. Object-based image analysis was applied and two classification methods were compared: NN (Nearest Neighbor) and CART (Classification and Regression Trees). The number of features was also examined to evaluate its effect on classification accuracy. The best accuracy is achieved by NN classifier with fewer features. When the classification results obtained through two datasets with NN method were compared, classification with more features showed lower accuracy. NN method with more features does not necessarily contribute to better classification accuracy. However, in the CART method, even if the number of features was varied, selected features were so similar that both accuracies were almost the same. It was suggested that it may be better to examine features carefully regarding its efficiency to separate each class and to sort out valid features in advance