森林立地学会誌 森林立地46(2), 2004, 93〜102

Jpn.J.For.Environment 46(2), 93-102 2004

 

Comparing the accuracy of predictive distribution models

for Fagus crenata forests in Japan

 

Tetsuya Matsui*, Tomoki Nakaya**, Tsutomu Yagihashi*, Hiroshi Taoda* Nobuyuki Tanaka*

*Forestry and Forest Products Research Institute

**Department of Geography, Ritsumeikan University

 

 We developed three different types of predictive distribution models for the presence/absence of Fagus crenata forests with four climatic parameters, and compared their performance. Generalised Linear Models (GLMs), Generalised Additive Models (GAMs), and Tree-Based Models (TMs) were compared in this way, due to their popularity in predicting plant species distributions. Four climatic factors; the minimum temperature of the coldest month (TMC), the warmth index (WI), winter precipitation (PRW), and summer precipitation (PRS); were used as explanatory variables in the model development. For GLMs, two sets of explanatory variables were applied, one of which was based solely on the four climatic terms (GLM-Simple), while the other included two-level interaction terms and quadratic polynomial terms (GLM-Complex). The models’ performance was compared with AIC (Akaike’s information criterion), residual deviance, and accuracy measures, including Kohen’s kappa statistic and overall prediction success, which are often used in predictive modelling studies. The resulting values all indicated that TMs performed best, followed by GAMs, GLM-Complex, and GLM-Simple. We envisaged that the superiority of the TMs may be due to their binary recursive partitioning nature, which appears to give them a high capacity to capture the non-homogeneous Japanese climatic patterns nationwide. The model can explain the relationships between F. crenata forest distribution and climate factors well, although the forests are widely distributed under non-homogeneous climatic systems in Japan. We therefore support the use of TMs in predicting the presence/absence distributions of widely distributed forest types or plant species under the Japanese climate systems.

 Key words: beech forest, empirical model, Generalised Additive Model, Generalised Linear Model, Tree-Based Model, climate

 

松井哲哉,中谷友樹,八木橋 勉,垰田 宏,田中伸行:ブナ林分布予測モデルの精度比較

ブナ林の分布予測モデル3種を気候値を説明変量として作成し,モデルの適合度を比較した。ブナ林の分布データは環境庁の3次メッシュ植生データからその有無を抽出し目的変量とした。気候データは気象庁の3次メッシュ気候値から抽出した月別気温・降水量データを基に最寒月最低気温(TMC),暖かさの指数(WI),夏期降水量(PRS)及び冬期降水量(PRW)の4気候値を計算して用いた。分布予測モデルは,植物分布解析でしばしば用いられる一般化線形モデル(GLMs),一般化加法モデル(GAMs),及びツリーモデル(TMs)の3種類のモデルを用いた。さらに,GLMsでは4気候変量のみを用いて作成した単純なモデルと(GLM-Simple),4気候変量にそれぞれの2乗項と2変数間の交互作用項を加えた複雑なモデル(GLM-Complex)を作成した。モデル精度の比較には,AIC(赤池情報量規準),尤離度,及び分布予測モデル研究でしばしば用いられるKappa統計量などの予測精度指標値を用いた。これらの指標値を比較した結果,TMsの適合度が高いことが判明した。TMsに続いてGAMsGLM-ComplexGLM-Simpleの順に適合度は高かった。TMsの適合度が高いのは,データをそれ以上分割しても無意味になるまで,かつ均質になるように2分割を続けていくことで,説明変数間の複雑な交互作用をモデル化できるTMsの特性が関係していると考えられた。すなわちTMsは,空間的に不均質な気候下の日本に広く分布するブナ林の分布をうまく説明できるモデルである。以上のことから,日本に広く分布する森林タイプや植物種の分布予測を行う場合の最適なモデルはTMsであるという結論に達した。

→Back to Home