종 분포 모형을 이용한 곰솔 잠재서식지 분포 예측 결과의 정확도 평가 연구 - 앙상블 방법론의 검증을 중심으로 -

Abstract

Species distribution models (SDMs) are widely used for biodiversity assessment, habitat management, and climate change impact assessment due to their ability to quantitatively evaluate species distribution. However, due to model uncertainty, the use of SDMs in public policy management has been limited. In order to overcome the limitations, many studies have been conducted mainly focusing on an ensemble approach, which compensates for the uncertainty of a single model. Even though ensemble methodology has been proven to improve accuracy compared to single models, this was based on inner validation. As inner validation has established flaws, with using the data in the form of ‘point‘, the need to assess outer validation with independent data in a polygon formations has been raised. In this study, we evaluated the accuracy of a Committee Averaging (CV) ensemble methodology using outer validation. In order to minimize uncertainty beyond the methodology setting, we used Pinus thunbergii, which has spatial specificity. As the outer validation method showed more accurate evaluation results, we used outer validation indices–sensitivity, specificity and accuracy–for comparison analysis between ensemble and single model results. Single models tend to overestimate compared to ensemble models, with a high value of sensitivity and a low value of specificity, whereas ensemble models tended to decrease the spatial uncertainty of single models, with generally high values of sensitivity, specificity and accuracy. Accordingly, the ensemble model methodology proved to improve accuracy by reducing the uncertainty of single models. Furthermore, through comparison analysis between outer and inner validation results, we additionally interpreted differences and limitations among inner validation, and have finally confirmed the need for further consideration in interpreting the results of the inner validation for both methodologies. Hence, outer validation using independent data should also be used.