高分解能衛星画像のテクスチャ特徴量とスペクトル特徴量を用いたオブジェクト指向型林分タイプ分類

抄録

This paper shows the object-based forest type classification using texture features from a panchromatic (PAN) image and spectral features from a multispectral (MS) image obtained by the QuickBird satellite.To investigate the performance of each feature, first, only texture feature is applied to image analysis, then spectral feature, and lastly combination of texture and spectral features.In this analysis, we use common segments obtained from a pansharpen image in order to compare the difference only between texture and spectral features.Distance between supervised classes is used to find well distinguishing feature combinations for classes.For PAN image analysis, 4 texture features from 8 candidates generated from co-occurrence matrix were selected. For MS image analysis, 9 spectral features from 10 candidates, such as 4 bands value and 6 differences between 2 bands from 4 bands, were selected.For PAN and MS analysis, 3 texture features from 8 candidates and all 10 spectral features were selected.Overall accuracy and Cohen's kappa of 6 forest types classification were 32.6% and 20.4% for PAN image, 74.6% and 70.6% for MS image, and 79.3% and 76.0% for PAN and MS images.This study demonstrated that combination of texture and spectral features exceeds a single feature in accuracy.