Soil salinization prediction for heterogeneous environments: an environmental similarity–based modeling framework

Soil salinization dynamics are driven by complex interactions among climatic conditions, hydrological processes, and anthropogenic activities. Due to this complexity, traditional single global models often struggle to capture spatial heterogeneity, leading to high prediction uncertainty and limited robustness at the pixel scale.

To address these challenges, this study proposes a multi-source data-driven framework based on environmental similarity matching to enhance prediction adaptability in heterogeneous environments. We compiled a dataset of approximately 35,000 topsoil samples from arid and semi-arid regions and constructed a multidimensional covariate system grounded in soil-forming factor theory. The framework comprises three components: (1) heterogeneity-based stratification, partitioning samples by climate and land use; (2) model library construction, developing candidate machine learning ensembles within each stratum via repeated cross-validation; and (3) similarity-based prediction, which employs Gower distance to quantify environmental similarity between target locations and training samples to select the optimal model.

Evaluations indicate that the Random Forest algorithm exhibits robust stability across stratified regions. Compared to single models, the environment similarity–constrained selection strategy significantly improved performance in heterogeneous regions; notably, the coefficient of determination (R²) in arid cropland areas increased from 0.748 to 0.807. Feature contribution analysis supports the necessity of stratified modeling, revealing that soil salinity in arid regions is primarily driven by vegetation variables and geographic, whereas remote sensing indices and soil pH dominate in semi-humid regions. The methodological framework developed in this study provides a new approach for high-precision soil salinity mapping.

KEYWORDS: Soil salinization; Environmental similarity; Heterogeneous environments; Machine learning.