Evaluation of the Factors Affecting Predictive Performance of Species Distribution Models
2013
Hochschulschrift
Zugriff:
101
Today’s forest policies and management strategies have been leading to forest conservation and public welfare with transference of society change and human life needs. For the sustainable use of bio-resources and maintain ecosystem function, how to tackle these operating problems arising in the management process and design an appropriate management principle to set nature reserve become an importance issues. The key to achieve the goal is that we need accurately describe species ecological and species geographic distribution information. Then we used species distribution model (SDM) coupled with statistical methods, remote sensing (RS), geographic information system (GIS) and global position system (GPS) to implement the procedure of analysis, combination, prediction and planning. Then SDM could exhibit species distribution pattern spatially by quantifying the relationship between species and environment. However, model performance is often affected by some factors such as species ecological characteristics, modeling algorithms, data quality, choice of predictor variables and so on. As mentioned above, prediction errors frequently occur in SDMs because these models usually simplify the real world while ignore important aspects of species ecology. Therefore, how to develop a robust model and interpret the effects of factors on model performance became an essential part of this study. Samples of Schima superba var. superba (Chinese guger tree, CGT) and Rhododendron formosanum (Red-stripe rhododendron, RSR) were obtained by GPS. GIS technique was used to overlay the layers of two species with topographic variables and spectral response variables. Species distribution models were developed by eight algorithms, including maximum entropy (MAXENT), DOMAIN, BIOCLIM, decision, tree (DT), discriminant analysis (DA), generalized linear model (GLM), maximum likelihood (ML) and back-propagation neural network (BPNN) to predict the potential habitats of the two species in Huisun study area, respectively. The study took split-sample validation approach and evaluated SDMs in terms of five indicators for model accuracy comparison. According data results indicated that four factors had significant effects on model prediction. Data quality had the highest influence, followed by species traits and model algorithm, and predictor variable selection was the lowest among them. All algorithms could not keep good performance across different training sample sizes, except MAXENT. The accuracies of eight models progressively increased with the number of samples until reaching a certain number, beyond which accuracies started leveling off and eventually reached maximum accuracy. Data resolution affected model’s accuracy, especially lower resolution data significantly degraded model performance. Moreover, the overall accuracy of RSR species was higher than that of CGT species. This means that species traits had more influence on SDM performance than algorithms. In terms of modeling techniques, MAXENT, DT and DOMAIN had the highest accuracy, followed by GLM, BPNN, and ML, and BIOCLIM, and DA had the lowest accuracy. Furthermore, the predictions of MAXENT, DT and DOMAIN generated high potential areas of CGTs and RSRs from the entire study area at the first stage, and thereby saving both cost and labor. More importantly, the models merely based on sample distribution and topographic variables could not be applied to predict species distribution at a larger spatial scale because the topographic attributes of Tong-Feng and Kuan-Dau watersheds are quite different from each other. Consequently, the prediction from all models with topographic variables built only from Tong-Feng samples could not be accurately extrapolated to the Kuan-Dau watershed. For a future study, direct factors (e.g. solar radiation, temperature and rainfall) or their surrogates and spectral response variables extracted from hyperspectral image data should be incorporated into an SDM so that it can be applied over a larger geographic area.
Titel: |
Evaluation of the Factors Affecting Predictive Performance of Species Distribution Models
|
---|---|
Autor/in / Beteiligte Person: | Chen, Hou-Chang ; 陳厚昌 |
Link: | |
Veröffentlichung: | 2013 |
Medientyp: | Hochschulschrift |
Sonstiges: |
|