Exploring on the climate regionalization of Qinling-Daba mountains based on Geodetector-SVM model

delimit comprehensive physical regionalization. Factor detector, one module of Geodetector, was used to explore factors that were important for different types of vegetation (with q-statis- tics greater than 0.3) and test their significance. The q-statistics formula is presented as fol- lows: q ¼ 1 Ns2 X L h¼1 Nhs2 h ð1Þ The study area was stratified into L strata, denoted by h = 1, 2, &, L. In this paper, h = 1 or 2, representing the subtropical zone and temperate zone. N and Nh are the number of samples in the total study area and the stratum h, respectively, σ2 and s2 h denotes the variation of sam- ples in the entire study area and the stratum h. A greater q value (the value of q ranges from 0 and 1) represents a stronger impact of a single factor on the spatial distribution of vegetation. The greater the q value, the stronger the relationship between the explained variable and the explanatory variables, and the larger the nonlinear association within them. In this paper, the explanatory variable are covariates (climatic factors, DEM and soil data), and the explained variables refer to representative vegetation type. For example, Pinus tabuliformis forest and Pinus massoniana forest, which represent the subtropical zone and the warm temperate zone, respectively. SVM. In this paper, the SVM model was used to delimit the subtropical and temperate zones according to the spatial distribution of typical vegetation. The SVM is a new approach developed in recent years and is a robust supervised learning algorithm [37]. Based on nonlin-

ear transformation, SVM maps the dataset to a higher dimensional space. It constructs a hyperplane between two classes that maximizes the margin between the closest training sam- ples of the classes to realize the classification of a dataset [48–50]. Specifically, the four raster layers of covariates represented by the index system are over- lapped into a multichannel dataset. The attributes are extracted from raster through the shape- file of the spatial distribution of representative vegetation. Grids with typical vegetation types were used as training sets. If subtropical vegetation is noted in the grid, the label of the grid is 0; if there is temperate vegetation in the grid, the label of the grid is 1. Then, each object of the shapefile has X and Y attributes. Here, X refers to temperature, precipitation, and DEM, whereas Y refers to the vegetation type. The training sets were used as input in the SVM model, and the result is output from the model. The result can be regarded as a zone map. Validation. Validation of the model included the following two aspects: boundary’s accu- racy and the analysis of the transitional characteristics. To verify the plausibility of the boundary, the plant community survey data and the previ- ous climatic boundaries were used for the validation. First, the plant community survey data were used as a reference for the real spatial distribution of vegetation. Specifically, we com- pared these data with the actual vegetation distribution to judge the degree of plausibility of

the results. In this paper, according to the “Flora Republicae Popularis Sinicae” and the plant community survey data [42], typical subtropical vegetation (Platycarya strobilacea forest and Pinus massoniana forest) and typical temperate vegetation (Pinus tabuliformis forest) were selected in this study for validation. In detail, Platycarya strobilacea and Pinus massoniana for- ests were used to validate the distribution of subtropical broad-leaved forests and subtropical coniferous forests. The plant community survey data with cell sizes of 20 m × 20 m were divided into three grades, including low (2-3), moderate (4-10), and high (11-98), according to the number of Platycarya strobilacea and Pinus massoniana forests. Similarly, Pinus tabulifor- mis forests were used to verify the distribution of typical temperate coniferous forests. The sur- vey data also were divided into three grades, low (2-10), moderate (11-20), and high (21-60), PLOS ONE Climate regionalization based on Geodetector-SVM model PLOS ONE | https://doi.org/10.1371/journal.pone.0241047 November 2, 2020 8 / 16