Abstract
Forest is a place of forestry activity. In addition, it has recently drawn public attention with its important function such as absorbing carbon dioxide, regulating its atmospheric concentration and subsequently affecting global warming, and maintaining high biodiversity. However, Japanese forestry is currently on the decline and, due to an increase in poorly controlled forests, the deterioration of these function is feared. Therefore, a way to efficiently manage forests with lower cost should be suggested for mitigating this situation. High-resolution remote sensing tools, which can collect detailed data for large areas, can be used effectively for this purpose. So, methods to extract forest data, such as forest density and tree heights, have been well examined. However, there has not been many researches carried out using these data and standard research methods using these data has not been established yet. Hence, it is significant to develop a new and effective method using these data for future research. In this study, we examined algorithm to classify forest types, and we classified forest into four types; natural conifer, natural broadleaved trees, artificial conifer and artificial broadleaved trees. The classification was carried out using spectral features obtained from IKONOS data, tree heights from LiDAR data, and texture features. The texture features used here were five variables that were the average, the contrast, dispersion, energy and entropy of R, G, B, Nir, NDVI. When the features were calculated, three rules were applied to each value of IKONOS and LiDAR data. The three rules were the followings. First, the section was between the maximum and the minimum in the histogram of the original data. Second, groups of cells (we call “window”) had four sizes; 3x3, 5x5, and 7x7. Final rule was that there were 2 types in window shape, namely rectangle and diamond shapes. The use of diamond shapes was a brand new method suggested in this study, and it was shown to improve the accuracy in the classification of natural conifer with this method. As a result of the analysis, the optimal size and shape of window, which gave the highest accuracy in the classification with texture features, were decided. The optimal shape was diamond-shape and the size was 7x7 divided into 7. Although the average classification accuracy of the whole area analyzed was only 61.3%, the accuracy of the natural conifer area was as high as 86.6%. This data used as much as 30 variables in Multi-Variate Analysis. However, even after decreasing the number of variables by Principal Component Analysis, the accuracy in the classification did not change largely. It was shown to be possible to decrease the number of variables in Multi-Variate Analysis without lowering the classification accuracy.
