利用羅吉斯迴歸與隨機森林預測臺灣杉造林地之變動──以林業試驗所六龜試驗林為例

Using Logistic Regression and Random Forest to Predict the Change of Taiwania cryptomerioides Plantations–An Example of Lioukuei Research Center in Taiwan Forest Research Institute

本研究利用1980及1990年航照正射影像，判釋並數化早期臺灣杉造林地邊界，以邊界為範圍套疊至2020年航照正射影像，從影像判釋造林地現況，並將現況林型區分為臺灣杉及非臺灣杉兩種林型，當做預測模型之因變量，另外選用高程、坡度、坡向、氣溫、雨量、林齡、距林道之距離及距造林地邊界之距離為自變量，利用羅吉斯迴歸與隨機森林預測早期臺灣杉造林地轉變為非臺灣杉林型之機率，透過預測模式準確度評估，羅吉斯迴歸分類精度表現為69.9 %，隨機森林分類精度可達82.4 %，整體而言，隨機森林較羅吉斯迴歸有極佳的預測效果，由兩種方法分析得到影響預測模型的共同重要變數為高程及離林道之距離。

This study uses the 1980 and 1990 aerial orthophotos in order to interpret and digitize early Taiwania cryptomerioides plantation boundaries. The forest types of Taiwania cryptomerioides and non-Taiwania cryptomerioides, being the dependent variables in the prediction model, are interpreted using 2020 aerial orthophotos. Besides, elevation, slope, aspect, temperature, rainfall, forest age, distance from roads, and distance from the border of the plantations are the eight factors being selected as independent variables. This study uses logistic regression and random forest with all of the above-mentioned variables in order to predict the probability of forest types changing from Taiwania cryptomerioides to non-Taiwania cryptomerioides. As a result, the classification accuracy of logistic regression is 69.9%, and the classification accuracy of random forest reached 82.4 %. Overall, the random forest has a better prediction effect than logistic regression. Based on the analysis results of both prediction models, the significant independent variables are elevation and distance from roads.