Examination of a classification model for damaged areas of windfall trees: Data analysis for damage caused by typhoon No. 23 in October 2004

The purpose of this study was to assess the suitability of a classification model for the determination of damaged areas of windfall trees caused by typhoon No. 23 in October 2004. Data input to the classification model encompassed seven factors: altitude, slope direction, slope angle, flow accumulation, curvature, wetness index, and land use type. The output data had four possible results: no damage and three levels of damage. A classifier was used for five classification algorithms. The eighteen modules included eight decision tree modules, two rule-based modules, two instance-based modules, two Bayes modules, and four function-type modules. Data reduction by clustering was used to improve the efficiency of our analysis by removing unnecessary data from the training data. The constructed models were evaluated using 'Sensitivity' and 'Specificity' indexes. The GIS analysis showed that large amounts of windfall-tree damage were observed in valley landform areas, which indicated that the soil moisture conditions such as flow accumulation and wetness index were high. Many of these areas were located on north, northeast and northwest facing slopes. As a result of modeling by a classifier, the instance-based algorithm outputted a relatively good classification. When performing data reduction by clustering, the classification accuracy tended to improve. When making estimates for an entire study area with a reduced number of samples, we found that it was possible to produce an adequate simulation for situations where the training data was reduced to around one-tenth of the original information, without negatively affecting the model performance. We suggest that the index of the multiplication of 'Sensitivity' and 'Specificity' is effective as an indicator of model performance.