Quantifying spatiotemporal dynamics of vegetation and its differentiation mechanism based on geographical detector

The influence of factors on vegetation changes in different regions is still largely unknown. We applied the geographic detector, a new spatial statistical method, to study the interactive effects of factors on the spatial patterns of normalised vegetation index (NDVI) changes and determine the optimal characteristics of key impact factors beneficial to vegetation growth. Our results show that from 2000 to 2015, the vegetation cover for the upper reaches of the Minjiang River, western China was in good condition. Furthermore, more than 80% of the areas had NDVI values ranging from 0.6 to 0.8 and NDVI > 0.8, and the spatial–temporal changes of vegetation cover were significant. The vegetation cover changes showed a significant transformation in the regions with NDVI > 0.6. Our study uniquely illustrated that elevation, annual average temperature and soil type can explain vegetation changes quite well. We propose that interactive effects exist among impact factors on vegetation NDVI, and the synergistic effects of the impact factors show mutual and nonlinear enhancements. The interactions among impact factors significantly enhance the impact of a single factor on vegetation changes. The most suitable characteristics of the main impact factors that promote vegetation growth were revealed by this study and will help improve our understanding of factors that impact NDVI and its driving mechanisms. Our findings suggest that the established favourable value range or the most suitable characteristics of impact factors will help management plans to intervene and promote vegetation change for vegetation restoration and alleviate environmental degradation.


Introduction
Surface vegetation is the main component of the terrestrial ecological environment that also plays a leading role in maintaining the functions of terrestrial ecosystems (Liu et al.2019a(Liu et al. , 2019b. Moreover, vegetation is a sensitive and vulnerable factor that reflects the changes in the ecological environment. Vegetation cover is an important remote sensing indicator for ecosystem monitoring and evaluation of surface processes and is also an important variable for understanding the ecosystem processes, health condition of vegetation and carbon storage (Luke Wallace et al. 2019;Ke et al. 2016). Therefore, the study of vegetation changes and their driving factors at regional and global scales is of great significance for monitoring the evolution of terrestrial ecosystems and changes in the ecological environment, managing natural resources to enable the development of strategies that adapt to environmental changes (Liu et al. 2013;Peng et al. 2019;Liu et al. 2019aLiu et al. , 2019b.

Responsible Editor: Philippe Garrigues
With the development of earth observation technologies, remote sensing-based vegetation indexes are frequently used for dynamic vegetation monitoring . The normalised vegetation index (NDVI) has been widely used in research regarding spatiotemporal scale changes in vegetation. Domestic and foreign scholars have recently applied remote sensing data and different research methods to study vegetation changes and their driving mechanisms. Piedallu et al. (2019) researched spatial variations in NDVI and their environmental drivers in mountains and Mediterranean biomes. Sanaei et al. (2019) quantified the effect of predictor variables on vegetation variations in the southern Alborz Province of Iran. Liu et al. (2019a) investigated the vegetation cover and its factors in the Wei River Basin, China. Mo et al. (2019) investigated the correlation among different temporal and spatial scales and strong effects of precipitation on NDVI in northwest China. Chen et al. (2019) analysed the vegetation cover and its correlation with climate factors in the Qinba mountains. Ma et al. (2018) analysed the grassland vegetation cover during the growing season in Qinghai Province and the influence of climate factors and human activities. Yao et al. (2019) studied the spatial heterogeneity of enhanced vegetation index vegetation and its relationship with vegetation greening in China. Sun et al. (2019) analysed the NDVI growth and its driving mechanism in the Luzangbo River Basin from 1982 to 2010. Liu et al. (2013) studied the changing trend in fractional vegetation cover (FVC) Yang et al. (2018) analysed the spatial patterns of FVC and its response to climate factors in Zoige area from 2000 to 2015. Xiong et al. (2018) studied the multi-scale coupling relationship between NDVI and environmental factors in the Maoxian region of the upper Minjiang River. Zhang et al. (2018) performed a quantitative analysis of vegetation coverage and its important impact on the upper Minjiang River Basin.
However, most of the above studies focused on the impact of climatic or natural factors on vegetation cover changes and establish their inter-relationship to explore and describe the changes and characteristics of vegetation cover. Although the findings are important and greatly promote the understanding of vegetation changes and their driving mechanisms, most of the methods involved in the studies applied linear, trend and correlation analyses. The researchers assumed a significant linear relationship between vegetation changes and their driving forces, which is not consistent with the actual situation . Simultaneously, with the development and popularisation of positioning and observation technologies, the problem of spatial heterogeneity was highlighted in either more elaborate or larger research studies on spatial big data; however, the statistical methods aimed for spatial heterogeneity are still very limited (Wang et al. 2017). Although hundreds of classification algorithms such as k-means and SOM have been used for classification or partitioning, statistical methods for spatial differentiation are still very limited (Wang et al. 2010. Currently, the main methods include geographic detector models for spatial heterogeneity measurement and factor analysis . Peng et al. (2019) quantified the natural factors responsible for vegetation NDVI changes in Sichuan Province, China. Zhu et al. (2020) quantified the effects of natural and anthropogenic factors on NDVI changes in the middle reaches of the Heihe River Basin. Hua et al. (2021) analysed the main drivers behind the spatiotemporal changes of desertification in the arid region of northwest China from 2003. Duan et al. (2020 identified the key factors influencing urban forest spatial differences within China. Geographic detector is a new tool for measuring, mining and utilising spatial heterogeneity, whose theoretical core involves detecting the consistency of the spatial distribution pattern between dependent and independent variables through spatial heterogeneity, and measuring the degree of explanation of the independent variables to dependent variables (Wang et al. 2017). The geographic detector model is stronger than general statistics, more confident and strongly suggests a causal relationship, because it is much more difficult for two variables to be uniformly distributed in two dimensions than for two variables to be uniformly distributed in one dimension.
The improvement in vegetation coverage in the upper reaches of the Minjiang River is of great significance for the construction of ecological barriers in the upper reaches of the Yangtze River Basin, supporting and serving regional and national survival and development, ecological civilisation construction, rural revitalisation and etc. Presently, less attention is paid on the way the changes in various impact factors and their superposition effects influence the changes in vegetation cover . Nevertheless, it is difficult to quantitatively analyse the factors influencing differences in NDVI variation . Although identifying the causes of vegetation cover changes is challenging, it is important to help us understand the connections among the multiple factors in the plateau environment in the upper reaches of the Minjiang River and the changes in vegetation cover. However, existing studies rarely discuss the relationship between vegetation cover changes and their multiple influencing factors. In addition to climatic factors, other natural factors (such as elevation, slope, aspect, soil and vegetation) and humanity factors (GDP and population density, distance from roads, rivers or town centres, etc.) also have a relatively great impact on vegetation cover changes. Instead of a single factor, the changes in vegetation cover are affected more by variations in multiple impact factors, especially in the upper reaches of the Minjiang River, where the topography is complex and multi-factor causes have complicated effects on vegetation cover changes.
The objectives of this study were as follows: (1) identify the main impact factors and their role in vegetation cover changes; (2) distinguish whether the impact factors for the upper reaches of the Minjiang River in Sichuan Province exert independent or interdependent influences on vegetation change; and (3) determine the optimum characteristics of the impact factors most conducive to vegetation growth. This paper comprises five parts: The "Introduction" section presents the introduction; the "Data sources and methods" section introduces the research area, data sources and research methods; the "Results and analysis" section reports the main results and findings of vegetation cover changes and influences on the factors driving NDVI change; the "Discussion" section presents a discussion on how impact factors within the optimum characteristics or value range serve to promote vegetation changes; the "Conclusions" section summarises the main quantitative results of the influence of impact factors.

Study area
The upper reaches of the Minjiang River, western China are located on the eastern edge of the Qinghai-Tibet Plateau and sited in the transition zone between the hilly lands surrounding the Sichuan Basin and the Qinghai-Tibet Plateau, spanning the coordinates 102° 34′-104° 14′ E, 30° 45′-33° 12′ N, which includes Songpan, Heishui, Mao County, Li County and Wenchuan, covering an area of 24,753.42 km 2 (Fig. 1). The geomorphic type of the upper reaches of the Minjiang River is dominated by plateaus and alpine valleys. The elevation rises from 762 m in the southeast to 5870 m in the northwest. The terrain is undulating, being high in the northwest and low in the southeast, with an average elevation of about 3400 m. The geological structure is complex, and the new tectonic activities are intense, causing several natural disasters such as earthquakes, landslides and debris flows. The climatic division of the upper reaches of the Minjiang River includes the mid-subtropical, northern subtropical and plateau climate zones. The climate varies from subtropical to temperate, frigid temperate to frigid, with obvious vertical differences. The reaches experience distinct dry and rainy seasons and uneven distribution of precipitation. The annual average precipitation is approximately 637.7 mm, and more than 80% of the rainfall is concentrated in May to October, with mostly strong and short-duration heavy rain. The temperature is relatively low, with small annual difference but large daily difference. The annual average temperature is 11 °C. The sunshine is sufficient, and the annual average land surface evaporation is 793.4 mm. The foehn effect in the valleys is significant whereby the annual rainfall is less than 500 mm, whereas the evaporation is 1340 mm. With the changes in elevation and hydrothermal conditions, the vegetation and soil types show remarkable vertical zonality. Vegetation types include forests, subalpine coniferous forests, subalpine meadows, alpine shrubs, arid valley shrubs and other ecological types. Soil types include cinnamon soil, brown soil, dark brown soil and subalpine meadow soil, all of which have typical vertical structures. In 2015, the population of the upper reaches of the Minjiang River was 39.17 × 10 4 people, and the total GDP was 148.73 × 10 8 yuan, in which the contribution of the first, second and third industries were 10.11%, 64.77% and 25.12%, respectively.

Data sources
We collected data for NDVI conditions and their driving factors from multiple sources. We used the annual maximum NDVI as a dependent variable to analyse the NDVI change and determine its mechanisms. The NDVI dataset from 2000 to 2015 was derived from the continuous time series of MODIS (MOD13Q1) using the maximum value composite method, and the sinusoidal projection of the MODIS (MOD13Q1) product was converted into a UTM projection; the coordinate system of the projection was set to WGS_84 based on the MODIS re-projection tool (MRT). We chose NDVI change value between 2000 and 2015 as the dependent variable, and selected representative and easily quantifiable 19 natural and human factors from the viewpoints of climate, topography, geomorphology, soil, vegetation and human activity with readily available data (Table 1). NDVI and DEM were derived from the United States Geological Survey (http:// gdex. cr. usgs. gov/ gdex/). GDP density, population density, climate, topography, vegetation, soil and Landsat 8 images were procured from the Data Center of Resources and Environmental Sciences of the Chinese Academy of Sciences (http:// www. resdc. cn). The land use data were derived by interpreting the Landsat 8 remote sensing images. The distance from county, township, road or river was obtained through GIS. GDP density and population density were calculated by applying the multi-factor weight distribution method. The GDP data or population with the administrative region as the basic statistical unit was distributed to the grid unit based on land use type, night light brightness, residential density and other factors closely related to human economic activities. Terrain data, that is, elevation, aspect and slope, were derived from the DEM. Climate map was extracted by interpolating the inverse distance weighted average method and DEM correction based on 1915 meteorological stations in China. Climate zoning data were compiled by the national meteorological administration in 1978, by using climatic data from 1951 to 1970. Soil map was compiled and published by the Chinese Soil Census Office in 1995. The vegetation map was digitised at the scale 1:10,000. The research data were uniformly set to the UTM, Zone 48 N, WGS_1984 projection coordinates and the GCS_WGS_1984 geographic coordinates.

Synthesis method of NDVI and its levels
Synthesis of NDVI The upper reaches of Minjiang River not only integrates multiple areas such as ecological barrier, biodiversity, Tibetan Plateau, typical fragile ecological environment and relatively concentrated poor population, but also is a complex system performing irreplaceable functions such as resource supply, ecological services and environmental regulation with economic and social development. To describe the characteristics of interannual NDVI more comprehensively, the Savitzky-Golay filter was used in this study to reconstruct the MODIS NDVI data to eliminate the influence of noise; the annual maximum NDVI was synthesised using the maximum value composite (

Image density segmentation and difference image algorithm
Setting the vegetation coverage threshold affects the calculation of vegetation coverage area. Considering the actual vegetation coverage in the upper reaches of Minjiang River identified using field survey data, and assuming that regions with a decrease in vegetation coverage in the range 0-10% may be omitted if the vegetation coverage threshold is set as 10% , dynamic changes in vegetation coverage were extracted using image density segmentation and image differencing algorithms. Vegetation coverage was assumed to either remain unchanged, decrease or increase if the difference values of vegetation coverage are zero, negative or positive respectively ).

Index selection and information extraction
Index selection The mountain system of the upper Minjiang is an extension of the Tibet Plateau, the formation of geomorphic types is complex, the diversity of climate, vegetation and soil, significant vertical gradient and strong disturbance owing to human activities has resulted in a complex and fragile ecological environment. According to the index system of selection of systematic, typicality, dynamic, scientific, quantifiable and can obtain the principle, only six categories and 19 factors including humanity factor, climate, geomorphic type, terrain, vegetation and soil data were selected to explore the influence of factors on vegetation changes in the upper Minjiang River (Table 1).
Information extraction A total of 24,709 random sampling points were generated based on 1 km × 1 km grids in GIS, and 23,161 valid samples were obtained after deleting invalid samples ( Fig. 1). Then, the NDVI and all factor data of sampling points were associated according to the spatial location to generate an attribute table, and the quantitative relationship between the corresponding NDVI and each index selection was calculated ).

Grading of impact factors
Based on the natural classification attributes inherent in the data, the similarity values of the classification interval are optimally grouped to maximise the difference between groups . Quantitative factors are classified according to natural breakpoints in the GIS (Liu et al. 2017), and non-quantitative factors are classified according to the number of their categories. Land use, geomorphic type, soil and vegetation were each classified into 6, 7, 18 and 8 classes according to their types, respectively. According to the natural break point method in the GIS (Liu et al. 2017), the GDP density, population density, distance from county, distance from town, distance from road, distance from river, annual average temperature, accumulated temperature (≥ 10℃), annual average precipitation, wetness index, total radiation, elevation and slope were each divided into 12 classes, aspect into 9 classes and dryness index into 6 classes, respectively (Table 1).

Variable coefficient
The variable coefficient of NDVI in the upper reaches of Minjiang River from 2000 to 2015 was calculated with the following formula: Here C v is the variable coefficient, n is the number of monitored years, F is the mean NDVI in study phase and F i is the NDVI in i year.

Geographical detector model
Geographical detector model represents a new spatial statistics method that is used to detect spatial heterogeneity and identify driving factors based on risk, factors, ecology and interaction (Wang et al. 2017;Wang et al. 2010). By calculating and comparing the q value of each single factor and the q value after the superposition of two factors, the geographic detector model can identify the presence of any interaction between the two factors and whether it is strong or weak, square, linear or nonlinear.

Detection of spatial heterogeneity and factors
The calculation method comprises the following steps: First, spatial overlay analysis was performed for the NDVI and factor layers; second, factors were divided into different spatial types or subzones; and third, a significance test for the differences of mean values of factors was conducted to detect the relative importance of the factors. The calculation model of the explanatory power of each factor is as follows: Here, q is the explanatory power of factors on vegetation NDVI, h = 1, …, L are the stratification of y or factor x, that is, classification or partition; Nh and N are the number of units in h and the whole region, respectively. N and σ2 are the total number of samples and the variance of y value in the whole region. Nh is the variance of units h.
The range of q value is [0, 1], and the larger the q value, the more obvious the spatial differentiation of y. In the extreme case, a q value of 1 indicates that factor x completely controls the spatial distribution of Y, and a q value of 0 indicates that the factor x has no effect on Y.
The variance calculation formula of the y value in the whole region is as follows: Here, Y j and Y are the values of the jth sample and the mean value of region Y in the study area, respectively. (1) Here, Yh,i and Y are the values of ith sample and the mean of Y in zone h, respectively.

Detection of factor interaction
Interaction detection is used to identify the interaction between factors, that is, to evaluate the accountability of the combined effect (enhancing or weakening) and respective effect on the NDVI. First, the q values of two factors with respect to NDVI were calculated (q(x i ) and q(x j )). Then, q values regarding the interaction between factors were calculated (q(x i ∩ x j )) and compared with q(xi) and q(xj).

Detection of risk zones
Risk detection is used to judge whether there is a significant difference in mean attribute values between the subzones of two factors, and can be used to find regions with high vegetation coverage. The risk detection is examined by using t statistic value:

Dynamic change in NDVI
In 2000, the regions of the upper reaches of the Minjiang River with middle (0.6 < NDVI < 0.8) and high vegetation cover (0.8 < NDVI < 1) areas accounted for 17% and 68%, 20% and 65% or higher in 2015 of the upper reaches of the Minjiang River, respectively. In contrast, the regions with low (0 < NDVI < 0.2), mid-low (0.2 < NDVI < 0.4) or middle vegetation cover (0.4 < NDVI < 0.6) area only accounted for less than 14% in 2000 and 2015 of the upper reaches of the Minjiang River (Fig. 2). This shows that the vegetation cover in the study area is overall satisfactory. However, the regions with 0.8 < NDVI < 1 reduced. From 2000 to 2015, the regions with 0 < NDVI < 0.2, 0.2 < NDVI < 0.4, 0.4 < NDVI < 0.6 and 0.6 < NDVI < 0.8 showed a rising trend, and proportion in the study area increased by 0.03%, 0.13%, 0.49% and 2.94%, respectively. The regions with 0.8 < NDVI < 1 showed a relatively large decreasing trend, with a reduction rate of 3.60% (Fig. 2).
The spatial distribution of vegetation cover showed obvious variation from 2000 to 2015. Overall, the regions with NDVI > 0.8 were mainly distributed as shrubs, meadows, broad-leaved forests and mixed coniferous broad-leaved forests at elevations below 3600 m in the upper reaches of the Minjiang River (Fig. 3). The areas with NDVI < 0.2 were mainly distributed in the extremely high-elevation regions in the western and north-eastern parts of the upper reaches of the Minjiang River at an elevation greater than 4500 m (Fig. 3). The areas with 0.2 < NDVI < 0. right). The areas with C v < 10% accounted for 79.02% of the total area, and were mainly distributed in areas with FVC > 0.8 and their margins. The area with 10% < C v < 60% accounted for 17.30%, mainly distributed on both sides in the southern, western and north-eastern regions of Minjiang River. The area with C v > 60% accounted for 3.68% of the total, mainly distributed in the very low vegetation coverage areas of high and extremely high mountains.
The changes in different classes of NDVI were calculated based on the NDVI spatial distribution statistics from 2000 to 2015 (Fig. 5). The NDVI showed a significant transformation in the regions with NDVI > 0.6, resulting in an increase in the areas with 0.6 < NDVI < 0.8 and a decrease in the areas with NDVI > 0.8 (Fig. 5). The roll-out areas with 0.6 < NDVI < 0.8 and NDVI > 0.8 were 1680 km 2 and 2080 km 2 , respectively, and the roll-in areas were 2415 km 2 and 1185 km 2 , respectively, which led to an increase and decrease in the areas with 0.6 < NDVI < 0.8 and NDVI > 0.8, respectively (Fig. 5). Vegetation cover changes were dominated by a rise in vegetation cover (0 < NDVI < 0.8) in the Minjiang River and its tributaries, showing a strip-like pattern. The vegetation cover in other areas increased and decreased in a canine-toothed pattern.

Influence analysis of detection factors
By calculating the q value of each impact factor, the influence exerted by each impact factor on NDVI was identified in this study (Table 2). All 19 factors exerted a significant effect on NDVI changes (p < 0.05). According to Table 2, the highest mean q values were observed for elevation, annual average temperature and soil type from 2000 to 2015, reaching 0.6006, 0.5261 and 0.3902 or higher, respectively, and the explanatory powers were 60%, 52% and 39% or higher, respectively. The mean q values for geomorphic type and annual precipitation were 0.1952 and 0.1950, respectively, and their explanatory powers were above 19%. The mean q values for dryness index, land use type, vegetation type, accumulated temperature (≥ 10 °C) and distance from road were 0.1649, 0.1546, 0.1447, 0.1383 and 0.1156, respectively, and their explanatory powers were all above 11%.
The q values of the remaining factors were all less than 0.09. Table 2 shows that the mean q values for annual precipitation, soil type and distance from county seat in 2008 were  (Table 2). Therefore, elevation, annual average temperature and soil type were identified as the important factors that influenced NDVI changes from 2000 to 2015 in our study area; annual precipitation and soil type were the main factors influencing the vegetation change in 2008.
From 2000 to 2015, the q value of other factors showed an increasing trend except for the decrease in the q value of land use, average annual precipitation and soil types. From 2000 to 2005, the q value of all other factors showed an increasing trend except for the decrease in the slope. From 2005 to 2015, the q value of other factors showed an increasing trend except for the decrease in the q value of GDP density and average annual precipitation. Except the increase in the q value of GDP density, population density, distance from county seat, humidity index, elevation and slope, the q value of other factors showed a decreasing trend during 2005-2008. From 2008 to 2010, the q value of factors showed an increasing trend except for population density, distance from the county seat, average annual precipitation, humidity index, global radiation, slope and aspect (Table 2).

Indication analysis of detection factors
The risk detector results can indicate how NDVI responded to changes in the level of a specific factor. The results showed that NDVI varied at all the levels of the different factors ( Fig. 6a-d), with acceptable statistical significance at the 95% confidence level (Table 3).
The larger the NDVI value of vegetation, the more suitable the features of each natural factor for vegetation growth. The difference in the mean vegetation NDVI value among different natural factors was significant (Table 3).
According to Fig. 6a, with the increase of GDP density, the mean value of NDVI fluctuated around 0.8507 and reached its maximum value (0.8507) at level 10 (77-104 yuan/km 2 ). With an increase in population density, the mean NDVI shows a relatively large fluctuation, reaching a maximum value of 0.8693 at level 8 (28.11-33.188 people/km 2 ) and followed by a downward trend (Fig. 6a). Within different land use factors, the mean NDVI shows a maximum of 0.8655 for grassland and woodland. The increase in the distance from the county first reduces the mean NDVI, shows fluctuation and then increases again. The maximum NDVI is 0.8490 at level 1 (0-7222.84 m) (Fig. 6a). For increasing distance from town, road or river, the mean NDVI shows a downward trend, with a maximum NDVI of 0.8617 at level 2 (3600.33-6322.39 km), 0.8610 at level 2 (1558.93-3410.16 km) and 0.8621 at level 2 (1576.01-3327.13 km), respectively (Fig. 6a). The mean NDVI showed an increasing trend for the annual average temperature and reached a maximum NDVI of 0.8843 at level 8 (6.3-7.95 °C) (Fig. 6b). According to Fig. 6b, for an increase in accumulated temperature (≥ 10 °C) and total radiation value, the mean NDVI shows a decreasing trend, and the maximum NDVI are 0.8878 at level 5 (15,643-20,428 °C) and 0.8644 at level 1 (3888.40-4042.48 MJ/m 2 ), respectively. With the increase in annual precipitation, the mean NDVI fluctuates, increases and then decreases, with a maximum NDVI of 0.8570 at level 6 (796.65-817.49 mm) (Fig. 5b). The mean NDVI  fluctuates with the increase in humidity index and reaches the maximum NDVI of 0.8140 at levels 8-15. With the increase in dryness index, the mean NDVI shows a fluctuating trend of increase, decrease, increase and then decrease, and the maximum NDVI is 0.8571 at level 2 (Fig. 6b). The mean NDVI shows a slight fluctuation and then decreases for different geomorphic types and elevations. The maximum NDVI values are 0.8045 and 0.8850 for hills or small relief mountains (low or middle mountains) and for those with a height of 2412-2748 m, respectively (Fig. 6c). The mean NDVI increases with the slope, showing a trend of increase and then decrease. The maximum NDVI is 0.8048 for a slope of 31.80-35.27° (Fig. 6c). For variations in aspect, the mean NDVI showed an upward-downwards fluctuating trend, reaching a maximum NDVI of 0.8058 between the range 292.5 and 337.5 (northwest slope) (Fig. 6c). The variation in mean NDVI shows different trends for different vegetation and soil types (Fig. 6c). The maximum NDVI of 0.8667 and 0.9390 were observed for broad-leaved and coniferous forests, and yellow brown, brown, dark brown and cinnamon soils (Fig. 6d).

Interaction analysis of detection factors
If the q value for factor interactions is greater than the maximum of both x i and x j but less than the sum of them (Max(A, B) < C < A + B), it indicates that the two factors are mutually enhanced. If the q value for factor interactions is greater than the sum of x i and x j (C > A + B), it indicates that the two factors are nonlinearly enhanced. The q values for interactions between most impact factors were higher than the q values of any individual impact factor, and the interactive influence of the impact factors is manifested as mutual enhancement effects and nonlinear enhancement effects (Table 4). The interactive influence of the geomorphic types, dryness index, annual precipitation, land use type, vegetation type, cumulative temperature and distance from the road with elevation shows mutually enhanced and nonlinear enhancement effects. The interactions among GDP density, annual average precipitation, total radiation, population density, distance from river, slope direction, slope, vegetation type and elevation show mutually enhanced and nonlinear enhancement effects. The interactions among the soil type, annual average precipitation, land use, geomorphic type, humidity index, global radiation, distance from township, river, road, vegetation type, cumulative temperature and annual average temperature showed mutual and nonlinear enhancement.

Major drivers of NDVI change
Elevation, annual average temperature and soil type have the greatest influence on NDVI changes among natural drivers. This is consistent with the findings of previous studies that vegetation is more sensitive to elevation Chen et al. 2019;Han et al. 2019;Liu et al. 2019aLiu et al. , 2019b, annual average temperature (Liu et al. 2013;Li et al. 2019;Vahagn et al. 2019) and soil type Liu et al. 2015) than other natural factors. We found an interesting pattern that precipitation gradients dictate the relative importance of environmental factors to vegetation in the study area. With an increase in elevation, the temperature drops, solar radiation and wind speed increase, precipitation and relative humidity first increase and then decrease in local regions and soil types show significant differences, causing changes in environmental gradients (Trujillo et al. 2012;Chen et al. 2019). This affects the vertical distribution and diversity of plants, and leads to different plant types and growth characteristics at different elevations (Liu et al. 2019a). Annual mean temperature dynamics lead to changes in other environmental factors (e.g. humidity and precipitation), producing a great superposition effect on plant growth and development (Liu et al. 2013;Vahagn et al. 2019). Related studies have proposed that soil type has a significant influence on vegetation growth by providing water and nutrients, and different soil properties lead to different vegetation distributions, which, in turn, protects the soil (Piedallu et al. 2019).

Effects of human activities
The mean q values of land use type and distance from the road were 0.1546 and 0.1156, respectively. The mean q values of the human factors, such as the distance from the river, global radiation, humidity index, population density, distance from county and GDP density,were all less than 0.09 (Table 2). Although the influence of the explanatory power of human factors was small, its combination with other factors can have a relatively large impact on the NDVI change. The impact of human on land use leads to the area changes of land use and mutual transformations between different land use types, which makes an important effect on vegetation cover. The impact of humans on land use leads to changes in land use areas and mutual transformations between different land use types, exerting an important effect on vegetation cover Wang et al. 2015;Han et al. 2019).

Effectiveness of the geographical detector model
Our research illustrated the effectiveness of the geographical detector model in detecting NDVI spatial differentiations and revealing their driving factors. Traditional principal component analysis, classical regression models and etc. usually rely on certain assumptions or constraints, such as normal distribution and linear hypothesis, to analyse the relationship between vegetation NDVI and its driving factors. Compared with the previous methods, the geographic detector is a new statistical method to detect spatial differentiation and reveal the driving factors behind it. This method has elegant form, clear physical meaning and no linear hypothesis. It uses spatial heterogeneity to detect the consistency of spatial distribution patterns between the dependent and independent variables and, based on them, measures the explanation power of the independent variables on the dependent ones. The method performs better in detecting the