【Objective】The speciation of metal elements in soil determines their environmental functions and effects. Developing predictive models for element speciation based on soil properties is an important approach to enrich the informational value of such data and reduce the number of required analytical indicators. This is of great significance for data mining under conditions of limited information. Most metal elemants, as important trace elements in soils, are widely present and affect crop growth and soil ecosystem health. Their forms and valence state have significant effects on their migration and transformation mechanisms on the surface and underground. Therefore, studying the metal forms in soil helps to understand their geochemical cycles and facilitate the evaluation of their impact on soil electronic networks, providing scientific basis for developing natural soil remediation methods and supporting the green, efficient, and sustainable use of soil. 【Method】This study selected manganese (Mn), a representative trace metal and redox-active element in soils, as the target. A total of 29 surface agricultural soil samples from different locations in the urban area of Hangzhou were collected and analyzed. The samples were characterized for their physicochemical properties, including total organic carbon (TOC), pH, total Mn content, and cation exchange capacity (CEC). The classical Tessier sequential extraction method was used to determine five Mn fractions in the soil: exchangeable, carbonate-bound, Fe–Mn oxide-bound, organic matter and sulfide-bound, and residual, and their correlation with soil physicochemical properties was evaluated. A neural network-based weight analysis method was then applied to predict Mn using soil physicochemical properties as input variables. 【Result】The results show that the soil pH was mainly alkaline, with abundant CEC and organic matter content. However, CEC exhibited high variability and was probably unevenly distributed and may be easily affected by external factors. Further analysis revealed that the average total Mn content in Hangzhou soils was 1.46 g·kg?1, higher than the background value for Zhejiang Province. Among the Mn fractions, Fe–Mn oxide-bound and residual forms were dominant, followed by organic/sulfide-bound, while exchangeable and carbonate-bound forms were the least abundant. Spatial distribution showed a layered pattern for exchangeable and carbonate-bound Mn, decreasing from north to south. Significant positive correlations were observed among most Mn fractions, except for the residual form. Among the physicochemical factors, pH showed the strongest correlation with Mn speciation, particularly a highly significant negative correlation with the exchangeable and carbonate-bound species. CEC was positively correlated with carbonate-bound and organic-bound Mn, while soil organic matter showed no significant correlation with any Mn fraction. Also, the neural network modeling demonstrated that using three parameters: total Mn, pH, and CEC, yielded the best prediction performance, with the coefficient of determination (R2) improving from 0.41 to 0.85, and prediction error reducing from 65% to 16%. 【Conclusion】The findings of this study provide theoretical support for predicting metal speciation in soils based on the observed distribution patterns of Mn and its relationships with soil physicochemical properties. The neural network-based modeling approach proposed herein offers a feasible strategy for deep mining of conventional soil survey data and enables rapid estimation of specific metal species. This contributes to a better understanding of the behavior of Mn in the soil redox network.