A study on the spatial assessment of site quality in the Manas River Basin based on the site index

The site index (SI) is a quantitative indicator of productive capacity at the plot level, whereas site quality represents its spatial evaluation. As a typical arid inland river basin, the Manas River Basin in Xinjiang requires reliable forest site quality assessment to support sustainable forest management under heterogeneous environmental conditions across this highly arid basin. This study focused on Spruce, Poplar, Elaeagnus angustifolia, Populus Euphrates Oliv, and Mixed Wood. Using the SI tables developed by our research team as reference data, environmental variables were extracted, and a site quality assessment model was established with a back-propagation (BP) neural network. By fitting the nonlinear relationships between environmental variables and SI, the model used plot-scale simulated SI values to predict site quality throughout the watershed and classified site quality levels through a dynamic classification method. The results showed that: (1) Pre was the core factor controlling site quality, whereas DEM exerted an indirect regulatory influence by affecting hydrothermal conditions; the BP neural network model showed satisfactory fitting performance for all studied tree species. (2) Site quality for each tree species displayed a distinct vertical zonation pattern: Spruce was concentrated in the southern mountainous region and had relatively high site quality; poplar occurred primarily in the central oasis plain, where site quality was moderate and uniform; Mixed Wood was distributed across the transition from the piedmont zone to the oasis, presenting the broadest distribution range and the strongest spatial heterogeneity; Elaeagnus angustifolia and Populus Euphrates Oliv were concentrated along the oasis margins and downstream desert belt, respectively, and both showed very low site quality. (3) Overall site quality across the basin was relatively low and exhibited a pronounced declining gradient from south to north. The "fair," "poor," and "very poor" classes together represented 82.6% of the study area, whereas the "excellent" class accounts for < 1%. From a methodological perspective, this study confirms the feasibility of combining SI tables with BP neural networks for continuous prediction of forest site quality, thereby reducing the spatial discontinuity associated with conventional plot-based surveys. Scientifically, a clear gradient from "high site quality–narrow ecological amplitude" to "low site quality–strong drought tolerance” was identified, providing the first spatially explicit assessment of species-specific habitat suitability within an arid inland river basin. Practically, the hierarchical management strategy developed from this assessment offer useful spatial information that, with field validation, can inform sustainable forest management in other arid regions along the northern foothills of the Tianshan Mountains. The findings provide a scientific basis for sustainable forest management in the Manas River Basin and in comparable arid regions located along the northern foothills of the Tianshan Mountains. However, given the moderate predictive performance for some species, these spatial results should be interpreted as indicative patterns requiring field validation.

Forest site quality assessment provides a scientific foundation for sustainable forest management and offers important technical support for the targeted improvement of forest resources. In 1971, the American Society of Forestry defined a site as “the environment of a forest’s growing area and the resulting vegetation type and quality on that land,” indicating that a forest site comprises the combined environmental conditions, including soil and climate, within a given area and its immediate vicinity (Rennolls et al., 1983). As a core indicator of forest productivity, the site index reflects the productive potential of a site through the height of dominant trees at a specified reference age and therefore has an irreplaceable role in forestry production practices (Wu et al., 2025a). In recent years, with the implementation of the “carbon peak and carbon neutrality” strategy, evaluating forest site quality has become increasingly important for guiding forest management and strengthening the carbon sequestration capacity of forest ecosystems.

Methods for evaluating site quality have developed from qualitative descriptions into quantitative models (Martínez-Salvador et al., 2025). A growing number of researchers have incorporated environmental variables into evaluations of site productivity; as early as 1967, soil and vegetation characteristics were used to assess the site quality of American pine plantations in Queensland, Australia (Pegg, 1967). To examine the effects of climate on site productivity, Monserud et al. (2006) were among the first to introduce climatic variables into forest site quality studies using pine stands as a case study. Traditional approaches generally construct SI curves from the relationship between dominant tree height and stand age. Holušová and Holuša (2025) reported that forest site conditions, including vegetation zone, elevation, slope aspect, and climate, determine biomass production potential, with standing volume differing by three- to eight-fold among site types. A productivity estimation model for forests of different ages was also developed using data from 2,048 permanent sample plots; this model explained more than 97% of the variation in aboveground biomass, emphasizing the importance of forest monitoring and improvement of growth models (Santiago-García et al., 2017).

With the integration of machine learning and remote sensing, new methods for forest site quality assessment have emerged. Huang et al. (2024) used BP neural networks, random forest, and the CatBoost algorithm to estimate forest carbon stocks. Their results showed that CatBoost achieved the strongest performance, with a test R 2 of 0.83, and indicated that window-based texture information was an important variable for constructing remote sensing estimation models. The “Afforestation Quality Index,” developed with a random forest model, progressively combines multi-source information, including natural vegetation characteristics, environmental variables, and climatic production potential (Zhang et al., 2024). Wu et al. (2025b) based on plot data from major coniferous forests across the country, developed site-class index models for 16 coniferous tree-species groups. All models achieved R 2 values above 0.95, while validation success rates exceeded 90%, providing a basis for evaluating forest site quality over broad spatial scales. For regional spatial mapping of site quality, another study used national forest inventory data and the random forest algorithm to establish a Composite Site Index (CSI) model. That study emphasized the need to account for spatial structure during model development in order to accurately map the geographic distribution of forest productivity (Subedi et al., 2024). The resulting model explained 81% of the observed variance.

In the development of SI tables, modeling approaches based on the relationship between diameter at breast height and tree height have received considerable attention. Previous studies have demonstrated that establishing a unified national model for site quality assessment is feasible; such an approach can reduce evaluation biases arising from climatic variation and differences in growth rates among forest age classes, thereby improving the cross-regional comparability of site quality. Although substantial progress has been achieved in forest site quality assessment, several limitations remain. First, most existing investigations have concentrated on humid and subhumid regions, whereas systematic evaluations of forest site quality in arid inland river basins remain relatively scarce (Yang and Wang, 2004). Second, the construction of conventional SI tables depends heavily on field-based plot surveys, making it difficult to adequately represent the spatial heterogeneity and continuity of site conditions. Nevertheless, the combined application of remote sensing data and machine learning technologies offers a new solution to this problem (Huang et al., 2024).

The Manas River Basin is situated at the northern foot of the Tianshan Mountains in Xinjiang, China. It represents a typical arid inland river basin and is also an important area for forestry and agricultural production in Xinjiang (Xu et al., 2025). Ecosystems within the watershed are relatively fragile, and forest vegetation performs an essential function in maintaining regional ecological stability and environmental security. Previous studies have shown that the ecosystem health index of the Manas River Basin displayed an overall increasing trend during the past 21 years, although it remained strongly constrained by climatic conditions and vegetation cover (Mu et al., 2025). Therefore, a systematic assessment of forest site quality in this region has substantial practical significance for guiding forest management and ecological restoration in arid environments. At present, two major problems remain in evaluating site quality across the Manas River Basin. First, plot-scale observations are insufficient for describing spatial continuity, and the regional variation patterns of forest site quality remain unclear (Zhang et al., 2021). Second, SI tables developed for the dominant tree species are lacking, which limits the availability of reliable reference standards for regional site quality evaluation.

To address these limitations, this study selected the Manas River Basin as the research area and focused on the following two objectives. First, based on the established SI tables and the BP neural network algorithm, a nonlinear mapping model between environmental variables and site indices was developed to extend site quality assessment from the “point” scale of individual sample plots to the “area” scale of the entire watershed. The hierarchical spatial distribution of site quality was then characterized using a dynamic classification method. Second, building on this spatially continuous assessment, the study clarified the vertical zonation patterns of site quality within a representative arid watershed. The resulting information can provide quantitative decision support for the rational allocation of tree species, regional forest restoration, and sustainable forest management throughout the basin.

Overview of the study area

The Manas River Basin in Xinjiang is located in the central part of the northern foothills of the Tianshan Mountains and along the southern margin of the Junggar Basin (Yu et al., 2010). Its geographic extent ranges from 85°01 ′ to 86°32 ′ E and from 43°27 ′ to 45°21′N. As the largest inland river basin along the northern foothills of the Tianshan Mountains, it covers a total area of approximately 19,800 km 2 (Cheng et al., 2002). An overview of the location and spatial extent of the study area is presented in Figure 1. From south to north, the basin can be divided into five major ecogeographic units: the southern mountainous region, the central oasis plain, the transition zone between the piedmont and oasis, the peripheral oasis zone, and the downstream desert region. Correspondingly, these units have formed relatively concentrated distribution areas for dominant tree species such as Spruce, Poplar, Elm, Chinese ash, Jujube, which differ in species composition and ecological adaptability (Shi et al., 2004). In the present study, elm, white ash, and honey locust had relatively few sampling plots and shared several similar biological characteristics. These species were therefore combined into a single category, referred to as Mixed Wood, for subsequent analysis. The Manas River Basin has a typical temperate continental arid climate, characterized by long and cold winters together with hot and dry summers. The long-term mean annual temperature ranges from approximately 6 °C–8 °C, while annual precipitation is generally about 150 mm–200 mm. In contrast, annual evaporation may reach 1,500 mm–2,000 mm, resulting in a comparatively high aridity index across the basin (Xu et al., 2004). Precipitation occurs mainly during summer, whereas seasonal snowmelt and rainfall jointly represent the principal sources of river recharge (Wang et al., 2011). These strongly uneven hydrothermal conditions contribute to substantial spatial differences in vegetation distribution, soil development, and forest growth potential. Soil types within the basin exhibit evident vertical zonation and horizontal differentiation. The piedmont plains are mainly characterized by light brown calcareous soils and gray desert soils, with textures dominated by sandy loam and silty clay loam and generally low organic matter content (Yu). Vegetation in the lower and more arid parts of the basin consists primarily of xerophytic and extremely xerophytic desert communities. The dominant plant communities include saxaul, Chinese broom, and tamarisk (Fan et al., 2010). These vegetation types have developed strong ecological adaptations to limited precipitation, high evaporation, and prolonged seasonal drought. During the past 50 years, increasing human activity has substantially altered the land-use structure and ecological environment of the watershed. Approximately 200,000 hm 2 of desert land within the basin has been reclaimed for agricultural production, transforming the region into an important cotton- and grain-producing base in Xinjiang (Zhong, 2020).

Data collection and processing

Sample data on the age and height of various tree species

The stand data used in this study were mainly obtained from the Xinjiang Forest Resources Survey, including records from the Xinjiang Forest Resources Continuous Inventory (hereafter referred to as the “Category I Inventory,” conducted every 5 years) and the 2011 Xinjiang Forest Resources Planning and Design Survey (hereafter referred to as the “Category II Survey,” conducted every 10 years). In 2011, the eighth Category I Inventory was implemented in the Xinjiang Uygur Autonomous Region, and tree measurement plots were established through a systematic sampling design. Topographic maps were divided into regular grids, with spacing of 4 km × 6 km in mountainous regions and 2 km × 3 km in plain areas. The permanent sample plots were square, measuring 28.28 m × 28.28 m and covering 0.08 ha. Category I inventory data combined systematic sampling with remote sensing techniques to provide a regional assessment of forest resources, including information on land type, stand age, forest area, and dominant tree species (Zan, 2016). Category II survey data supplied more detailed stand-level information, mainly including stand age, diameter at breast height, and dominant tree species. Using the age and height records for each tree species organized by classification code, together with the 2000–2020 Xinjiang forest age dataset developed by our team (Zhai et al., 2025), we extracted the corresponding spatial coordinates for the distribution of dominant tree species in 2011. Data preprocessing included replacing missing observations with the median value of the corresponding variable, converting continuous aspect angles into categorical aspect classes, and removing sampling points lacking key variables. These procedures prepared the dataset for subsequent analyses intended to extend site index estimates from the plot scale to the watershed scale. The complete dataset was then used for model development following the procedures described in Section 2.3.3.

Weather data

The DEM data used in this study were obtained from the Earth Big Data Science Data Center of the Chinese Academy of Sciences1 at a spatial resolution of 30 m. Slope and aspect were derived from the DEM using ArcGIS surface analysis tools. Precipitation (Pre), temperature (Tmp), and potential evapotranspiration (Pet) data were acquired from the National Qinghai-Tibet Plateau Scientific Data Center,2 and the original products were provided at a monthlyscale. To satisfy the annual matching requirements of the site quality assessment, all meteorological variables were aggregated to the annual scale for 2011. The year 2011 was selected as the reference year because the corresponding forest inventory data were relatively complete, spatially continuous, and temporally consistent. The DEM and meteorological datasets were uniformly resampled to a 30 m spatial resolution, after which the extracted raster layers were stacked to construct the environmental feature dataset for site quality assessment.

SI data

The SI data used in this study were obtained from SI tables developed by our team for five dominant tree species (Spruce, Populus, Elaeagnus angustifolia, Populus euphratica Oliv, and Mixed Wood) in the Manas River Basin. Following established procedures, five candidate growth equations—the Richards, Gompertz, Korf, Schumacher, and Weibull functions—were fitted to identify the optimal guide curve for each species. The reference age of each species was determined by jointly considering the coefficient of variation in dominant tree height and the corresponding growth pattern. SI class intervals were subsequently defined using the relative dominant height method (Kitikidou et al., 2015,; Luo et al., 2019). The resulting SI tables were evaluated using an independent point-check validation procedure, and the validation success rate exceeded 90% for all five species groups. The complete tables include the fitted guide curves, selected reference ages, SI class boundaries, and validation results. The specific SI tables and test results are provided in Appendix 1. These tables were developed specifically for the five dominant tree-species groups in the Manas River Basin using locally collected age–height observations. Although the resulting tables are region-specific, the overall methodological framework—including candidate growth-equation fitting, reference-age selection, SI class determination, and independent validation—can be transferred to other tree species and regions when sufficient local age–height data are available for calibration.

Research methods

The theoretical basis for site suitability index evaluation

Site quality represents the potential productivity of a forest and is jointly determined by multiple environmental conditions, including climate, topography, soil, and hydrology; however, because it is an abstract ecological concept, it cannot be measured directly. The SI is the most widely applied quantitative proxy for evaluating site quality (Wang et al., 2024). It is generally defined as the dominant height attained by a specified dominant tree species at a reference age and provides an intuitive indication of the relative productivity of a particular site.

A close but nonlinear relationship exists between SI and underlying site quality. Sites with similar environmental quality may produce different SI values because of differences in tree-species composition, stand structure, or human management, whereas comparable SI values may arise from different combinations of environmental factors (Wei H. et al., 2024). Thus, SI can be treated as an “observable proxy” for site quality, while a spatial evaluation of site quality should further identify how environmental variables regulate its geographic variation.

Accordingly, this study used plot-level SI values as the dependent variable and DEM, slope, aspect, and Pre as independent variables to establish a nonlinear mapping model, thereby extending site quality estimates from individual sample plots to the entire watershed (Dong et al., 2024). Because the association between environmental variables and SI is complex and strongly nonlinear, conventional statistical approaches may not characterize it sufficiently. Therefore, a BP neural network was selected for model construction because it has strong nonlinear mapping ability, is relatively insensitive to multicollinearity, and does not require a predefined functional form. The BP neural network captures the complex nonlinear correspondence between environmental factors and SI values obtained from the SI tables and subsequently derives a general mapping function (Liang et al., 2025). The mathematical formulation is provided in Equation 1. By applying the fitted function to every raster cell throughout the watershed, an SI value can be estimated for each grid location. The predicted SI value is used to represent site quality as defined in this study; after classification, it is converted into a site quality grade, thereby providing methodological support for continuous spatial assessment of forest site quality.

In this equation, X i represents the input environmental factor; W ij and W j are weights; b j and b out are thresholds; N h is the number of hidden nodes; f hid is the logsig activation function [ f (z) = 1/(1 + e z)]; and f out is the purelin linear activation function [ f (z) = z ].

Screening for environmental factors

The Pearson correlation coefficient (Sun et al., 2008) was used to analyze the correlation between each environmental factor (DEM, Slope, Aspect, Pet, Pre, Tmp) and SI, in order to identify the key factors that significantly influence site quality (Ding et al., 2004). The Equation 2 for calculating the correlation coefficient is:

Here, x i represents the environmental factor (e.g., DEM, slope, etc.) at the i -th sampling point, y i represents the SI at the i -th sampling point, n represents the total number of sampling points, and r represents the correlation coefficient between the environmental factor and the SI.

Variance inflation factors (VIF) were used to diagnose multicollinearity among the environmental factors. A VIF value < 10 was considered to indicate no severe multicollinearity, and all factors were retained for subsequent modeling. The Equation 3 for calculating VIF is:

Here, VIF j is the variance inflation factor for the i -th independent variable, where j is the index of the independent variable; R j 2 is the coefficient of determination for the regression of the j -th independent variable on all other independent variables.

Based on the results of the correlation analysis and multicollinearity diagnostics, the environmental factors ultimately included in the model were identified, thereby providing a basis for selecting the input variables for the BP neural network model.

Based on the multicollinearity diagnostics, four environmental factors—DEM, Slope, Aspect, and Pre—were selected as input variables. DEM was retained because it integrates regional hydrothermal variation through its correlations with both precipitation and temperature. Slope and Aspect were included as they influence surface runoff, erosion, and solar radiation receipt, all of which affect local site conditions. Tmp and PET were excluded due to strong collinearity with Pre (see Section 3.1 for the correlation results), which would otherwise inflate parameter variance and increase overfitting risk.

Methods for building BP neural network models

In recent years, machine-learning approaches, including random forest, have been increasingly used for forest site quality assessment. Compared with such ensemble methods, BP neural networks provide three advantages for the present study. First, they generate a continuous and differentiable mapping function that can be integrated into a spatial prediction framework. Second, they can perform well with moderately sized datasets without extensive hyperparameter optimization. Third, their architecture can be explicitly defined to represent the assumed ecological structure, with an input layer for environmental drivers, a hidden layer for nonlinear transformation, and an output layer for SI prediction. This structural clarity is consistent with our aim of developing a spatially explicit relationship between environmental conditions and forest site quality. Therefore, the BP neural network was selected because it provides a balance among predictive performance, interpretability, and suitability for the available dataset.

The basic topology of a BP neural network comprises an input layer, a hidden layer, and an output layer. The input layer transfers environmental signals to the hidden layer; the hidden layer processes these signals according to the connection weights and specified activation rules, and then passes the transformed information to the output layer; finally, the output layer integrates the processed signals and produces the predicted result. During supervised learning, the model output is compared with the observed value to calculate the prediction error. This error is then propagated backward through the network to update the connection weights, thereby completing the iterative learning process (Liu, 2017). As shown in Figure 2, the network contains an input layer with four nodes corresponding to DEM, Slope, Aspect, and Pre, one hidden layer with k nodes, and an output layer with a single SI node. The hidden layer adopts the logsig activation function, whereas the output layer applies the purelin linear activation function. The selected number of hidden nodes (k) differs among tree species: five nodes were used for Spruce and Elaeagnus angustifolia, and 2 for Poplar, Populus Euphrates Oliv, and Mixed Wood. Solid arrows represent forward propagation, whereas dashed arrows denote backward propagation of the prediction error.

Following the data preprocessing steps described in Section 2.2.1, we constructed BP neural network models to establish the nonlinear mapping between environmental variables and SI. Before training, the input variables, including DEM, Slope, Aspect, and Pre, were transformed by Z-score standardization (Equation 6) to reduce the influence of differing units and numerical scales. Meanwhile, the output variable SI was rescaled to the [0, 1] interval through min–max normalization (Equation 4) to match the output range of the purelin activation function (Cheng, 2023).

The complete dataset was subsequently divided into training and testing subsets at a ratio of 8:2. However, because the sample plots are spatially distributed, this random split may overestimate predictive performance due to spatial autocorrelation; the resulting R 2 values should therefore be interpreted as model fitting performance rather than fully independent prediction accuracy. Using DEM, Slope, Aspect, and Pre as input variables and SI as the output variable, a single-hidden-layer BP neural network was developed with the "neuralnet" package in R. The selected learning algorithm was resilient backpropagation with weight backtracking (rprop+), and a linear activation function was assigned to the output layer. The appropriate number of hidden-layer nodes was identified through a trial-and-error procedure. For each candidate architecture containing two to five hidden nodes, constrained by the available training sample size, the model was trained three times. The optimal node number was selected according to the highest mean test-set R 2 and a training success rate above 30%. Finally, the chosen network architecture was retrained three times using the complete training dataset, and the run with the best performance was retained as the final model. Selection of the hidden-layer size was also supported by the empirical relationship presented in Equation 5.

Here, x scaled refers to the standardized value of the variable, x refers to the raw value of the variable (such as raw observations of environmental factors like DEM or slope), x min refers to the minimum value of the variable, and x max refers to the maximum value of the variable.

In this equation, N h represents the number of hidden layer nodes; N in represents the number of input layer nodes; N out represents the number of output layer nodes; and h is an integer selected from the range of 1–10.

In the equation, X Z represents the transformed value, x represents the value before transformation, μ is the mean of all sample data, and σ is the standard deviation of all sample data.

The optimal network architectures determined for each tree species are presented in Table 1. To accelerate network convergence during training, the standardized input variables were further truncated to the interval [−3, 3], thereby limiting the influence of extreme outliers. The main modeling parameters were specified as follows: target accuracy = 0.001, learning rate = 0.01, and maximum number of iterations = 1,000. The hidden layer employed the logsig activation function, whereas the output layer adopted the purelin linear function (Table 2). A complete description of the model settings, including activation functions, learning rate, maximum iterations, convergence threshold, and training algorithm, is provided in Supplementary Table 2.

Site quality assessment and classification

Based on the statistical distribution of the predicted SI values, the mean and standard deviation were used as reference measures, and μ ± 1.5 σ and μ ± 0.5 σ were selected as threshold points for dividing site quality into five grades. The detailed classification standards are presented in Table 3. This approach determines grade boundaries according to the distributional properties of the predicted data, thereby reducing the subjectivity introduced by manually defined classification thresholds. It can objectively characterize the relative distribution of forest site quality and has been widely adopted in previous site quality assessment studies (Huang et al., 2006; Teng et al., 2009; Shen et al., 2018). This classification framework is also consistent with the spatial evaluation of site quality for individual tree species and forest types conducted in the present study.

In this study, the coefficient of determination (R 2) and root mean square error (RMSE) were used as metrics to evaluate model performance (Dong et al., 2022). The resulting performance metrics for both training and testing datasets, along with the corresponding sample sizes, are reported in Table 4. The Equations 7 and 8 is as follows:

Here, y i represents the observed value, ŷ represents the predicted value, ȳ represents the mean of the observed values, and n represents the sample size.

Development and accuracy validation of site quality evaluation models for different tree species

To identify key environmental variables shaping site quality, we first quantified correlations between each candidate predictor and species-specific SI values. As specifically shown in Figure 3. Of the four topographic and climatic variables retained for analysis, Pre had the strongest absolute correlation with SI for four of the five tree species: Spruce (r = 0.26), Poplar (r = −0.35), Elaeagnus angustifolia (r = 0.01), Populus euphratica Oliv (r = −0.09), and Mixed Wood (r = −0.20). For Elaeagnus angustifolia, SI was most strongly associated with DEM instead of Pre. Before formal variable selection, Pre was the climatic variable most closely correlated with SI; Tmp and potential evapotranspiration (PET) were discarded owing to severe collinearity with Pre (| r | > 0.7). We thus retained Pre as the only climatic predictor alongside DEM, Slope and Aspect. Overall weak linear correlations imply that SI cannot be adequately explained by any single environmental factor, justifying the use of a nonlinear multi-factor framework (i.e., BP neural network) to disentangle complex environmental controls over site quality.

Using these four predictor variables, separate BP neural network models were constructed for each tree species. Model performance was evaluated using the coefficient of R 2 and RMSE for both the training and testing datasets, with the results summarized in Figure 4. Across the five species, the test-set R 2 values ranged from −0.19 to 0.37. Mixed Wood obtained the highest test R 2, followed by Poplar, whereas spruce exhibited a moderate level of predictive performance. By contrast, Elaeagnus angustifolia and Populus euphratica Oliv. showed relatively weak generalization ability, with test R 2 values of −0.06 and −0.19, respectively. The generally moderate or low R 2 values, especially for the two drought-adapted species, may primarily result from their limited sample sizes and the intrinsic complexity of forest growth, which is controlled by many ecological and management factors beyond the four environmental predictors considered in this study. Despite these constraints, the BP neural network still provides a feasible method for the spatial evaluation of forest site quality in arid inland river basins.

Spatial distribution characteristics of site quality classes for plots of different tree species

Based on the spatial distribution maps and proportional composition of site quality grades for Spruce, Poplar, Elaeagnus angustifolia, Populus Euphrates Oliv and Mixed Wood presented in Figure 5, clear differences among species can be observed. Spruce is mainly concentrated in the southern mid-mountain zone, where moisture availability is comparatively favorable. Its site quality is primarily classified as “good” or “fair,” reflecting the relatively strong dependence of Spruce on moist and suitable mountainous habitats. Poplar is mainly distributed across oasis plains and river corridors, where site quality is generally rated as “fair” or “good” and shows a comparatively uniform spatial pattern. Elaeagnus angustifolia occurs mainly within the oases and along their margins in the middle and lower reaches of the river; its site quality is predominantly categorized as “poor” or “fair,” which may restrict its overall growth potential to some extent. Populus Euphrates Oliv, in contrast, is distributed almost exclusively along downstream desert riverbanks, where site quality is overwhelmingly classified as “poor” or “very poor.” Mixed Wood is scattered throughout the transitional region between the piedmont and oasis zones; its site conditions are generally fair, and its relatively dispersed distribution demonstrates the capacity of mixed-species combinations to buffer environmental variation and adapt to transitional habitats. Overall, spruce forests in the mountainous region are mainly associated with relatively high-quality sites, whereas Poplar and Mixed Wood constitute the principal forest types in the oasis and transitional zones. Elaeagnus angustifolia and Populus Euphrates Oliv function as important ecological stabilizers under more extreme site conditions. This spatial arrangement provides a clear geographic foundation for accurately estimating forest carbon sinks in the Manas River Basin and for developing differentiated forest management and ecological restoration strategies.

Among the five tree species groups in the Manas River Basin, clear differences were also observed in the proportional distribution of site quality grades. Among stands categorized as “very poor,” spruce accounted for 94.1% of its total distribution area, indicating a considerable presence on the lowest-quality sites and suggesting strong tolerance to poor soil and drought conditions. Mixed Wood occurring under very poor site conditions represented 60.1% of its total area and was likewise mainly distributed in relatively harsh habitats. Elaeagnus angustifolia and Populus Euphrates Oliv were also concentrated on lower-quality sites, with poor and very poor classes accounting for 59.1% and 45.2% of their respective total areas. Among these two species, Elaeagnus angustifolia appeared to show slightly broader adaptability than Populus Euphrates Oliv. Poplar exhibited the highest requirement for favorable site conditions, with approximately 39% of its distribution occurring on excellent sites and only 5.9% occurring on very poor sites. Overall, Spruce, Mixed Wood, Elaeagnus angustifolia, and Populus Euphrates Oliv are the predominant vegetation types on relatively low-quality sites within the watershed, whereas poplar was largely restricted to higher-quality ecological environments. Future ecological planning should therefore be adapted to local site conditions: spruce and Mixed Wood may be prioritized in areas with very poor conditions, Elaeagnus angustifolia and Populus Euphrates Oliv are more suitable for areas with moderate-to-poor conditions; and poplar cultivation should be concentrated in locations with relatively favorable site quality.

For Elaeagnus angustifolia and Populus euphratica Oliv, the sparse distribution and limited spatial extent—restricted to the oasis margins and downstream desert riparian zones—make their spatial patterns less visible at the basin scale. See Table 4 for details. This limited distribution reflects their ecological specialization to extreme arid environments. Quantitative summaries of area percentages for each site quality class are provided in Table 4 to supplement the visual interpretation for these two species.

Spatial distribution patterns of site quality for different tree species

By extending the analysis from individual sample points to a continuous spatial scale, this study identified the geographic distribution of site quality for each tree species. As illustrated in Figure 6, the spatial patterns of site quality differed clearly among the five tree-species groups. Spruce site quality ranged from 15.230 to 25.751, with a pronounced high-value zone concentrated in contiguous areas of the southern Tianshan Mountains and forming continuous belts along the mountain slopes. This distribution was generally consistent with the precipitation pattern on the northern slope of the Tianshan Mountains. Low-value areas appeared mainly within the piedmont transition zone, further reflecting the sensitivity of Spruce growth to moisture availability and topographic conditions. Poplar site quality ranged from 14.465 to 16.849, indicating a relatively high overall productivity level. Poplar was continuously distributed across the central oasis plains, forming broad and relatively connected areas. Under the influence of the oasis irrigation system, its site quality showed a comparatively uniform spatial distribution and stable growth potential, suggesting that Poplar is a suitable tree species for establishing artificial shelterbelts. The site quality of Elaeagnus angustifolia ranged from 7.063 to 7.465 and remained extremely low overall. This species was mainly distributed within the oases and along their margins in the middle and lower reaches, where it exhibited a fragmented and patch-like spatial pattern. As a drought- and salt-tolerant species, Elaeagnus angustifolia can persist under severe environmental conditions; however, its productive potential was strongly restricted, and the observed site quality approached the lower limit of its physiological tolerance. The SI of Populus Euphrates Oliv ranged from 8.450 to 8.693, with its distribution concentrated mainly within the downstream desert riparian zone, where narrow and elongated belts followed ancient channels and modern river courses. Mixed Wood site quality ranged from 5.049 to 11.227, representing the broadest numerical range and the highest degree of spatial heterogeneity among the investigated forest types. These forests occurred as scattered patches throughout the transition zone between the piedmont and oasis areas. Their relatively diverse tree species composition allowed adaptation to a broad range of site conditions, extending from poor to above-average quality, and reflected the typical compositional characteristics of forests in ecologically heterogeneous transition zones.

A comprehensive comparison among the five tree-species groups showed that Spruce had the highest site quality, the widest variation in SI, and pronounced spatial heterogeneity. It occurred in continuous belts throughout the southern mountainous region, responded strongly to differences in environmental conditions, and therefore represented the most sensitive indicator species of site quality within the watershed. Poplar also exhibited relatively high overall site quality, but its SI range was narrow and its spatial heterogeneity was limited. It formed continuous, patch-like distributions across the central oasis plains and was strongly influenced by the homogenizing effect of irrigation systems. Both Elaeagnus angustifolia and Populus Euphrates Oliv belonged to the low-site-quality group, with low mean values, narrow SI ranges, and very limited spatial variability. Elaeagnus angustifolia occurred in fragmented patches along oasis margins in the middle and lower reaches, while Populus Euphrates Oliv formed narrow, elongated belts within the downstream desert zone. Both species displayed an apparent “plateauing” pattern in site quality, suggesting that their growth potential was strongly suppressed by prolonged exposure to extreme environmental stress. Mixed Wood had moderate- to low-overall site quality but showed a comparatively broad range and strong spatial heterogeneity. It was distributed in scattered patches across the transition from the piedmont to the oasis and was the only forest type covering site conditions from low to moderate or relatively high quality, demonstrating the complex habitat adaptability of species combinations in this ecological transition zone. Overall, the five forest types— Spruce to Populus Euphrates Oliv —displayed a sit quality differentiation pattern of “highest → relatively high → extremely low → extremely low → slightly below average” across the Manas River Basin. Spatial heterogeneity generally showed an inverse relationship with distributional continuity, providing a clear geographic basis for implementing differentiated forest management strategies throughout the basin.

The area statistics presented in Table 4 further support the spatial distribution characteristics shown in Figure 5, particularly for sparsely distributed species that are difficult to display clearly on basin-scale maps.

A comprehensive study on site quality for various tree species

After determining the site quality values and corresponding quality grades for the five selected tree species groups, an integrated spatial assessment of forest site quality across the Manas River Basin was conducted, as presented in Figure 7. Site quality within the basin displayed an evident vertical zonation pattern. In the southern mountainous areas of the upper reaches, the “good” and “excellent” grades were dominant, generally with SI values greater than 24; this region was therefore classified as having high site quality and was considered suitable for the growth of species such as Spruce and Poplar. In the foothill zone of the middle reaches, the “Fair” grade occupied the largest proportion, forming a moderate-quality site region that was relatively suitable for Mixed Wood. By contrast, the northern plains and downstream desert areas were dominated by “poor” and “very poor” grades, with generally low SI values below 18, and were consequently classified as low-quality site regions. These areas are mainly suitable for drought-tolerant tree species such as Elaeagnus angustifolia and Populus Euphrates Oliv. Overall, site conditions across the watershed exhibited a pronounced decreasing gradient from south to north. This spatial pattern provides a clear basis for decision-making related to forest management, rational tree species allocation, and ecological restoration throughout the basin.

At the basin scale, overall site quality was relatively low, with the “fair,” “poor,” and “very poor” categories jointly accounting for 82.6% of the total area. In comparison, the “good” and “excellent” categories represented only 16.5% and 0.9%, respectively, and the area classified as “excellent” was < 1%. The spatial distribution was also strongly unbalanced, with clear contrasts between high-quality and low-quality regions. These findings indicate that land with favorable conditions for efficient forestry and agricultural use is highly limited within the Manas River Basin. Future ecological development should therefore give priority to protecting the remaining high-quality sites, while differentiated measures, including site improvement, vegetation conservation, and ecological restoration, should be applied in areas classified as “fair,” “poor,” and “very poor.”

An explanation of the mechanisms by which environmental factors affect site quality

This study identified ‘Pre’ as the principal environmental factor controlling forest site quality in the Manas River Basin. Pre exhibited a positive relationship with the SI values of all investigated tree species, which is consistent with the widely recognized ecological pattern that water availability represents the primary limitation on vegetation growth and spatial distribution in arid and semiarid regions (Achite et al., 2017; Wan et al., 2023). Pre directly affects site productivity by supplying plant-available water, while its spatial variation, largely shaped by mountain-related orographic effects, establishes the fundamental environmental framework of forest site quality across the basin. Thus, the regional gradient in water availability appears to be the main factor underlying the observed differences in forest growth potential.

However, the strength of the relationship between Pre and SI differed considerably among tree species, revealing a clear ecological sensitivity gradient from moisture-responsive species to drought-adapted species. Spruce and Poplar exhibited the greatest sensitivity to Pre, reflecting their predominant distribution in the southern mid-mountain areas and central oasis plains, where water conditions are comparatively favorable. This result agrees with earlier findings showing that orographic uplift in the Tianshan Mountains increases precipitation by approximately 10–20 mm for every 100 m increase in elevation (Lu, 2009). In contrast, Elaeagnus angustifolia and Populus euphratica Oliv displayed much weaker responses to Pre because these species can persist in downstream desert regions receiving <150 mm of annual precipitation by relying on groundwater resources and strong drought and salinity tolerance. Mixed Wood showed an intermediate level of precipitation sensitivity and exhibited the widest range of site quality together with the greatest spatial heterogeneity. This pattern may reflect the buffering effect of tree-species diversity, which reduces dependence on a single environmental factor and improves adaptability to heterogeneous ecological conditions (Jie et al., 2007; Yu et al., 2023).

Topographic variables, particularly DEM, influenced site quality indirectly by regulating regional water and thermal conditions. DEM was strongly positively correlated with Pre and strongly negatively associated with temperature and potential evapotranspiration, indicating that elevation can serve as an integrated indicator of hydrothermal variation in arid environments. Higher elevations generally receive more precipitation and experience lower temperatures and evaporative demand, whereas lower areas are characterized by reduced rainfall and stronger water loss. Because severe multicollinearity was detected among Pre, Tmp, and potential evapotranspiration, only Pre was retained as the climatic predictor to reduce the risk of model overfitting and avoid inflation of parameter estimation (Shi et al., 2013). Accordingly, the final modeling framework included four environmental predictors: DEM, Slope, Aspect, and Pre. The resulting ecological sensitivity gradient—from moisture-responsive (Spruce, Poplar) through intermediately responsive (Mixed Wood) to drought-adapted (Elaeagnus angustifolia, Populus euphratica Oliv)—provides a mechanistic explanation for the spatial patterns of site quality described in the following sections.

Ecological interpretation of spatial patterns in site quality

As shown in Figure 8, which illustrates the statistical distribution of SI values, the site quality of different tree species in the Manas River Basin displays a distinct vertical zonation pattern. The SI of spruce ranges from 14.5 to 25.8, with a median value of approximately 20. Among the five investigated tree-species groups, spruce has the highest overall SI and the broadest numerical range, with a positively skewed distribution concentrated mainly in the southern mid-mountain zone. This pattern indicates its strong dependence on relatively moist habitats and reflects the narrow ecological amplitude of the species (Chen et al., 2012). The SI range of Poplar was 15.2–16.8, with a median of approximately 16. Its values are generally high but show only limited variation, resulting in a narrow and relatively uniform distribution. This pattern likely reflects the compensatory influence of the oasis irrigation system on natural precipitation limitations, together with the relatively restricted ecological range of poplar (Wei L. Z. et al., 2024). The SI of Mixed Wood ranges from 5.0 to 11.2 and exhibits considerable internal variation and strong spatial heterogeneity, covering site conditions from low to upper-middle quality. This broad variation reflects the capacity of tree-species diversity to buffer environmental stress and adapt to the complex habitat conditions of the piedmont–oasis transition zone (Zeng et al., 2010); The SI ranges of Elaeagnus angustifolia and Populus Euphrates Oliv are 7.1–7.5 and 8.5–8.7, respectively. Both species remain at very low levels and display extremely narrow ranges, indicating that their growth potential is strongly constrained by drought, salinity, and poor soil conditions in the downstream desert zone. These limited SI variations further suggest that both species may be approaching their physiological tolerance thresholds under the severe envirnmental conditions of the lower watershed (Liu, 2017).

To determine the spatial suitability of each tree species in the Manas River Basin, this study first predicted the spatial distribution of site quality classes for each species and subsequently delineated the boundaries of their suitable habitats. According to the dynamic classification results presented in Table 3, grid cells classified as “Good” (Class II) or “Excellent” (Class I) for a particular species were extracted from the complete set of classified raster cells and defined as suitable habitat for that species (Holušová and Holuša, 2025). The specific suitable distribution ranges of the five tree-species groups are presented in Figure 9. As a species with a relatively narrow ecological range, Spruce has strict moisture requirements and is primarily restricted to the southern mid-mountain zone, where precipitation is comparatively abundant and site quality reaches its highest level. Although poplar also has relatively high water requirements, artificial water supplementation from the oasis irrigation system allows it to occur continuously and uniformly across the central oasis plains, where site quality is moderate to high and spatial heterogeneity remains limited. Mixed Wood is mainly distributed throughout the transition zone between the piedmont and oasis areas. Owing to the complementary ecological characteristics of its component tree species, this forest type can occupy a relatively broad range of site conditions, extending from poor to above-average quality, and consequently exhibits substantial variation and strong spatial heterogeneity; Elaeagnus angustifolia and Populus Euphrates Oliv adapt to extremely arid environments through different adaptive mechanisms: Elaeagnus angustifolia relies largely on its tolerance to salinity, drought, and nutrient-poor soils, while Populus Euphrates Oliv utilizes groundwater through an extensive root system. Both species are concentrated in the downstream desert region, where predicted site quality is extremely low and SI values show an evident plateauing pattern. Their productive capacity has therefore been compressed toward the lower limits of physiological tolerance, even though they remain capable of maintaining important ecological functions under harsh environmental conditions.

In summary, the site quality distributions of the five tree species groups in the Manas River Basin form a clear ecological gradient. In the southern mountainous region, spruce is characterized by high site quality, a narrow ecological range, and strong sensitivity to moisture availability. This pattern gradually shifts toward the central oasis plains, where poplar predominates under moderate-to-high site quality, depends partly on artificial irrigation, and displays a relatively homogeneous spatial distribution. The gradient then extends into the piedmont–oasis transition zone, where Mixed Wood is characterized by moderate site quality, a wider ecological amplitude, and a stronger capacity to buffer environmental heterogeneity. Finally, in the downstream desert, where species such as Elaeagnus angustifolia and Populus Euphrates Oliv —characterized by extremely low site quality because of their pronounced adaptation to drought, salinity, and limited water availability. This ecological sequence is broadly consistent with elevation-related vegetation gradients reported for mountain ecosystems in China and other regions (Whittaker and Niering, 1975,; Fontaine et al., 2007). It reflects the combined influences of climatic differentiation, topographic control (Wang et al., 2019), species-specific physiological adaptation (Li, 2008), and human intervention through irrigation and land management (Chen and Wei, 2022), thereby providing a clear spatial basis for forest management and ecological restoration within the watershed.

Previous evaluations of site quality in semiarid regions have demonstrated that differentiated management is an effective approach for improving the efficiency of ecological restoration (Lai et al., 2018). Areas with high site quality should receive priority protection to maintain their existing productivity and ecological functions. In regions with moderate or low site quality, tree species should be selected according to local environmental conditions and their ecological adaptability. For areas with extremely poor site quality, management should emphasize ecological restoration, vegetation stability, and the prevention of further degradation rather than high productivity. The observed distribution patterns also indicate that the five tree-species groups may differ in their vulnerability to future climate change. Moisture-sensitive Spruce and Poplar may experience habitat contraction or declining productivity under continued warming and increasing evaporative demand, particularly where water supplementation is unavailable. By contrast, drought-adapted Elaeagnus angustifolia and Populus euphratica Oliv may tolerate increasing aridity more effectively, although their growth potential is already strongly limited under current site conditions (Huang et al., 2014; Liu et al., 2021). Understanding these species-specific sensitivities and adaptive capacities is therefore essential for developing climate-responsive forest management and restoration strategies in the Manas River Basin.

Analysis of the applicability of BP neural network models and limitations of the research

In arid inland river basins, tree growth is influenced by the combined effects of highly heterogeneous environmental and biological factors, including precipitation, elevation, soil moisture, salinity, and interspecific competition; therefore, even a BP neural network cannot completely predict site quality for every species when only a limited number of environmental variables are available. Nevertheless, compared with conventional multiple linear regression, the BP neural network still provides several clear advantages (Wang et al., 2021). First, it does not require the prior specification of a functional relationship and can adaptively identify complex nonlinear patterns within the observed data (Wu et al., 2012). Second, it is relatively insensitive to multicollinearity among predictor variables, which makes it suitable for ecological datasets containing strongly correlated environmental factors (Liang et al., 2025; Wang et al., 2021).

However, several limitations of the present study should be recognized. Important environmental variables, including soil organic matter, soil salinity, groundwater depth, and distance from rivers, were not incorporated into the model. The absence of these factors may be particularly important for drought-adapted species such as Elaeagnus angustifolia and Populus euphratica Oliv, whose growth is probably controlled more strongly by groundwater availability, salt accumulation, and proximity to river channels than by regional precipitation alone (Yao et al., 2012). In addition, the relatively sparse distribution of sample plots may increase spatial interpolation uncertainty, especially in downstream desert regions where forest stands are fragmented. The “black-box” characteristics of BP neural networks also make it difficult to quantify the direction and magnitude of individual environmental contributions. Moreover, random division of the dataset into training and testing subsets may overestimate model performance because nearby plots can share similar environmental conditions due to spatial autocorrelation. Consequently, the reported R 2 values may partly represent spatial interpolation ability rather than fully independent prediction accuracy. The use of static environmental and forest inventory data from 2011 also prevents the model from describing temporal changes in site quality caused by climate variation, land-use change, irrigation development, or forest degradation. Furthermore, no systematic comparison with alternative algorithms was conducted, which limits the evaluation of whether the BP neural network was the most suitable modeling approach. These limitations should be considered when interpreting the spatial prediction results and applying them to management decisions.

Future research should address these limitations by incorporating soil and hydrological variables, including groundwater depth, soil moisture, soil salinity, soil organic matter, and distance to permanent river channels (Xu et al., 2020). Long-term remote sensing time-series data should also be integrated to characterize temporal changes in vegetation condition, forest productivity, and site quality. Spatial cross-validation methods are needed to reduce the influence of spatial autocorrelation and provide a more realistic assessment of model generalization. In addition, interpretable machine-learning approaches, such as SHAP could be used to quantify the relative contribution and interaction effects of different environmental factors. Comparative evaluations involving random forest, gradient boosting, support vector regression, and other algorithms would further clarify the relative performance of BP neural networks. Finally, the proposed framework should be extended to other arid inland river basins along the northern foothills of the Tianshan Mountains to evaluate its regional transferability and ecological applicability.

This study integrated species-specific SI tables with BP neural networks to attempt spatially continuous prediction of site quality for five dominant tree-species groups in the Manas River Basin, extending the assessment from discrete sample plots to the watershed scale. Because the predictive accuracy of the models varied across species, the following conclusions should be viewed as hypothesis-generating patterns that require further field verification. Given the moderate model performance, the main findings are summarized as follows:

(1) Environmental drivers and model performance

Pre was identified as the principal climatic factor influencing forest site quality within the watershed, while elevation exerted an indirect effect by regulating regional water and heat conditions. BP neural network models using DEM, Slope, Aspect, and Pre as environmental inputs captured the nonlinear relationships between these variables and plot-level SI values. Model performance differed among tree species, with moderate predictive ability for some species and weaker generalization for species represented by limited samples or strongly affected by unmeasured hydrological factors. Despite this variation, the results suggest the potential of applying BP neural networks to spatial site quality assessment in arid inland river basins.

(2) Vertical zonation of tree-species site quality

Site quality followed a clear south-to-north ecological gradient. Spruce was concentrated in the southern mountainous region and generally occupied the highest-quality sites. Poplar was mainly distributed across the central oasis plains, where site quality was relatively high, spatially uniform, and influenced by artificial irrigation. Mixed Wood occurred principally within the piedmont–oasis transition zone and displayed the widest SI range and strongest spatial heterogeneity. By contrast, Elaeagnus angustifolia and Populus euphratica Oliv were generally predicted to be restricted to the oasis margins and downstream desert with extremely low quality.

(3) Overall basin assessment and differentiated management strategies

Overall forest site quality in the Manas River Basin was relatively low and showed a pronounced decreasing gradient from south-to-north, while areas classified as “excellent” accounted for < 1% of the basin. Based on these spatial differences, a tentative tiered management strategy is proposed. High-quality forest sites should receive priority protection and limited disturbance; moderate-quality areas should be managed through appropriate species selection, stand improvement, and efficient water-resource allocation; and poor or very poor sites should focus mainly on ecological restoration, soil and water conservation, and the maintenance of drought-tolerant vegetation. Under projected climate warming and increasing aridity in Central Asia, these differentiated strategies should be periodically revised using long-term monitoring data and updated climate projections to ensure their effectiveness under changing environmental conditions.

In summary, this study makes three key contributions. Methodologically, it demonstrates a novel integration of SI tables with BP neural networks to produce spatially continuous predictions of forest site quality. Scientifically, it quantifies a clear high-to-low site quality gradient and reveals its species-specific ecological implications across mountainous, oasis, transitional, and desert environments. Practically, it establishes a differentiated management framework for forest conservation, species allocation, and ecological restoration in arid regions. These findings, pending field validation, can inform sustainable forest management in the Manas River Basin and in comparable arid inland river basins along the northern foothills of the Tianshan Mountains.