土壤有机碳（SOC）是陆地生态系统最大的碳库，在复杂地形下表现出巨大的空间异质性，严重影响土壤碳库的估算。微地形变化如海拔、坡度、坡向等都对土壤有机碳进行空间异质性的调配起着重要作用。本研究融合地学统计分析与先进的机器学习技术，结合微地形及土壤特性进行模拟研究，旨在提高复杂地形下土壤碳库的估算精度。

研究基于广东省黑石顶自然保护区的已建立的 50 公顷森林大样地

（1000m×500m）区域内 837 个采样点的土壤碳库、微地形和其他土壤理化性质数据。通过莫兰指数分析，发现 SOC 呈现出显著的空间自相关。经 Box-cox变换的SOC 数据通过普通克里金插值显示其半变异函数符合指数模型，函数块金值为 1.1×10-1,偏基台值为 2.22×10-1，变程为 474m；并且得到了黑石顶森林SOC 的等级分布。随后，采用随机森林模型，对选定研究区的地表特征进行了深入分析。研究分别构建了以微地形特征以及微地形与土壤特性为变量的两个模型，并通过网格搜索和五折交叉验证法进行训练和模型测试。结果显示，考虑土壤理化性质的过程影响后，将微地形对土壤碳库空间变异的解释度从

0.33 提升至 0.59，RMSE 从 8.78 降至 6.93；微地形中的海拔因子和土壤特性中的有效氮含量是影响SOC 预测最重要的指标。微地形与土壤特性模型的性能优于仅考虑微地形特征的模型，具有更好的适应性和建模效果。

本研究不仅强调了考虑土壤特性在地表特征预测中的重要性，也为环境和资源管理提供了决策支持。未来研究可探索多源数据融合、应用高级机器学习技术及扩展到更广阔的应用领域，以进一步挖掘地学数据的潜力同时有助于提高SOC 的预估精度，并可为其他相关的科学研究提供参考。

Soil organic carbon （SOC）, the largest carbon pool in terrestrial ecosystems, shows great spatial heterogeneity under complex topography, which seriously affects the estimation of soil carbon pools. Microtopographic variations such as elevation, slope, and slope orientation all play an important role in the deployment of soil organic carbon for spatial heterogeneity. In this study, we integrated geostatistical analysis and advanced machine learning techniques, combined with microtopography and soil properties to conduct simulation studies, aiming to improve the estimation accuracy of soil carbon pools under complex terrain.

The study was based on soil carbon pools, microtopography and other soil physicochemical property data from837sampling points in an established 50 ha forested large sample plot （1000m×500m） area in Heishiteng Nature Reserve, Guangdong Province. The SOC showed significant spatial autocorrelation as analysed by Moran's index. The Box-cox transformed SOC data showed that its semi-variance function conformed to the exponential model by ordinary kriging interpolation, with a function nugget value of 1.1×10-1, a skewed abutment value of 2.22×10-1, and a variance range of 474m; and the class distribution of SOC in the Black Rock Top Forest was obtained. Subsequently, the surface features of the selected study area were analysed in depth using the random forest model. The study constructed two models with microtopographic features and microtopographic and soil properties as variables, respectively, and trained and tested the models by grid search and five-fold cross-validation method. The results showed that considering the process effects of soil physicochemical properties increased the degree of explanation of spatial variability of soil carbon pools by microtopography from 0.33 to 0.59, and decreased the RMSE from 8.78to6.93; the elevation factor in microtopography and the effective nitrogen content in soil properties were the most important indicators affecting the prediction of SOC. The microtopography and soil characteristics model outperformed the model considering only microtopographic features, with better adaptability and modelling effectiveness.

This study provides decision support for environmental and resource management. Future research could explore multi-source data fusion, application of advanced machine learning techniques and expansion to broader application areas to further exploit the potential of geomorphological data while contributing to the improvement of SOC prediction accuracy.