Driving force analysis of soil quality across China through an explainable deep learning framework

Soil is crucial for human survival and sustainable development. However, research on large-scale soil quality assessment and its variability mechanism remains largely unexplored. This study introduces a novel Property-Function-Quality (PFQ) framework to evaluate soil functions and quality, and an explainable deep learning approach, the Convolutional Neural Network-based Integrated Gradients (CNN-IG) model, to decipher the spatial variation of soil quality by quantifying the individual and interactive effects of driving factors. Applied across China, the PFQ and CNN-IG approaches revealed that: (1) the Soil Quality Index (SQI) ranged from 0.133 to 0.890 across the country, with higher values in the Qinghai-Tibet Plateau and lower values in the northern windy and sandy regions; (2) the CNN model outperformed traditional Artificial Neural Network (ANN) models in soil functions and SQI simulations, with determination coefficients ( R 2 ) 5.4–21.3% higher than those of the ANN model; (3) climate factor contributed most significantly to the spatial variation of SQI nationwide (60.5%), followed by geographical (28.6%) and human factors (10.9%); and (4) the influence of individual factors on SQI was significantly shaped by interactions with other factors. This study provides a novel and interpretable framework for large-scale soil quality assessment and mechanistic insight. The findings emphasize the dominant role of climate in shaping soil quality and highlight the importance of factor interactions. These results offer valuable guidance for soil conservation and sustainable land management under changing climatic and anthropogenic scenarios.