EMULATING THE HABITAT OF DIFFERENT TREE SPECIES AT LOW-MEDIUM ELEVATION BY MACHINE LEARNING

摘要

Numerous studies have applied a geospatial information system (GIS) coupled with machine learning algorithms to build species distribution model (SDM) for simulating the habitat of rare plant species with narrower ecological amplitude (EA) on a specific characteristic (e.g. elevation), but seldom for species with broader EA since simulating the latter is generally more difficult than the former. This study selected Randaishan cinnamon (Randaishan cinnamon, RC) and Rhododendron formosanum (Formosan rhododendrons, FR) as targets with different EAs for comparison. Moreover, data for terrain-related variables (e.g. slope) were used in this study since data for them can be easily acquired by remote sensing. DEMs of three grid sizes (5, 20, and 40 m) were used to derive these variables, including elevation, slope, aspect, terrain position (TP), surface curvature (SC), profile curvature (PRC), and plan curvature (PLC). The SDMs were generated using discriminant analysis (DA), decision tree (DT), random forest (RF). and support vector machine (SVM) that were developed using the machine learning module (scikit-learn) written in Python programming language. Both values of the Kappa coefficient of agreement and Matthews correlation coefficient (MCC) for RF are the best, DT is the second, and SVM and DA are the worst among them. Regardless of which algorithm being used, the accuracies of FR models are at least 10% greater than those of RC models. Of more importance, not only it is more efficient to simulate steno-species (i.e. FR) than eury-species (i.e. RC), but also the predictive ability of the former is superior to that of the latter. FR species has much stronger adaptability to poor circumstances with thin, infertile, and acidic soils than RC species does. Furthermore, it has been found that the TP (i.e. slope position from ridge or peak to valley) plays a vital role, followed by elevation, and other variables vary with the resolution and algorithm. DEMs and their derived terrain-related variables can improve the accuracy of SDM substantially as the spatial resolution raises from 40m to 5m, and there are more variable combinations to achieve the highest accuracy.