Objective Rapid urbanization in arid regions presents distinctive challenges for air quality management, particularly concerning fine particulate matter (PM_2.5). This study aims to systematically quantify the seasonal dynamics of PM_2.5 concentrations across different local climate zone (LCZ) types within a high-density arid city. It seeks to elucidate how two-dimensional landscape patterns and three-dimensional urban morphological characteristics jointly influence the spatial distribution of PM_2.5, and to identify the dominant drivers and their nonlinear mechanisms in this unique climatic context.

Methods The main urban area of Ürümqi, a representative high-density city in the arid northwest of China, was selected as the case study. A multi-source data fusion framework was condtructed, integrating satellite remote sensing data (Sentinel-2 and Landsat 8/9 imagery), vector-based architectural data, ground-based meteorological observations, and high-resolution topographic data. Methodologically, the study proceeded in two main stages within the overarching LCZ framework. First, a random forest (RF) model was employed to generate high-resolution (30-meter) seasonal PM_2.5 concentration maps through inversion techniques and to perform a precise LCZ classification for the study area. Second, an eXtreme Gradient Boosting (XGBoost) machine learning regression model, coupled with the SHapley Additive exPlanations (SHAP) interpretability framework, was applied. This advanced analytical approach was used to deconvolve the relative importance and, more importantly, the nonlinear dependence and threshold effects of a comprehensive set of influencing factors. These factors encompassed two-dimensional landscape metrics, three-dimensional urban morphological indicators, elevation, and key meteorological parameters.

Results The analysis revealed a pronounced seasonal pattern of “higher PM_2.5 concentrations in winter and lower in summer” in Ürümqi’s main urban area, coupled with a spatial distribution characterized by “higher concentrations in the north than in the south, and in built-up areas compared to green spaces”. Significant differences in PM_2.5 levels were observed among LCZ types. LCZ 10 (heavy industry) and the compact built types (LCZ 2, compact midrise and LCZ 3, compact low-rise) were identified as persistent high-pollution zones. In contrast, forested LCZ types (LCZ A, dense trees and LCZ B, scattered trees) exhibited a significant capacity to mitigate PM_2.5, maintaining consistently low concentrations. Factor importance analysis indicated seasonal shifts in the dominant controls. The NDVI emerged as the most influential factor in summer, demonstrating a linear negative correlation with PM_2.5. A threshold effect was observed, with NDVI values greater than 0.25 leading to a marked enhancement of its purifying effect during both seasons. In winter, air temperature and elevation (digital elevation model-DEM) became the predominant factors. Temperatures below −10.2 °C strongly favored the formation of temperature inversions, trapping pollutants near the surface. Concurrently, areas with elevations below 800 m, particularly in the northern basin, were prone to forming “cold-air pools” that exacerbated pollution accumulation. Other key nonlinear thresholds were identified: a bare land cohesion index (COHESION_BLG) exceeding 85 in winter led to a sharp increase in dust emission potential; an open-set LCZ cohesion index (COHESION_open) greater than 88 facilitated better pollutant dispersion; and a temperature above 25°C in summer promoted vertical mixing and PM_2.5 diffusion. Furthermore, the compact LCZ group consistently showed significantly higher pollution levels than the open-set LCZ group, highlighting the role of urban morphology in modulating air quality. SHAP analysis further quantified several other key nonlinear thresholds: a Bare Soil/Sand group cohesion index (COHESION_BLG) exceeding 85 in winter led to a sharp increase in dust emission potential; an open-built-type LCZ cohesion index (COHESION_open) greater than 88 facilitated better pollutant dispersion; and a temperature above 25 °C in summer promoted vertical atmospheric mixing and PM_2.5 dispersion. Furthermore, the compact LCZ group (LCZ 1−3) consistently exhibited significantly higher pollution levels than the open-set LCZ group (LCZ 4−6), unequivocally highlighting the decisive role of urban morphology compactness in modulating local air quality.

Conclusion This study provides a comprehensive and quantitative analysis of the complex interplay between multi-dimensional urban morphology and PM_2.5 pollution in an arid, high-density city, leveraging the standardized LCZ framework. It successfully advances the application of the LCZ scheme in arid-region air pollution research, moving beyond qualitative associations to delineate clear seasonal divergences in underlying controlling mechanisms. The principal contribution lies in the innovative integration of explainable machine learning (specifically, XGBoost-SHAP), which enabled precise quantification of critical nonlinear thresholds of key morphological, topographic, and meteorological factors. These findings transcend a merely deeper mechanistic understanding. The findings yield concrete, quantitative scientific evidence that can directly inform the development of precise, LCZ-type-specific and seasonally-adapted PM_2.5 management strategies. Consequently, this study offers a robust, evidence-based foundation for optimizing urban spatial planning and urban design in Ürümqi and other arid-region cities that face similar air quality challenges.