基于机器学习方法的宁夏南部土壤质地空间分布研究

【目的】基于历史数据,利用机器学习方法分析宁夏南部土壤质地空间变异规律及其与环境因素之间的关系。【方法】基于宁夏回族自治区南部428 个20世纪80年代第二次土壤普查土壤剖面点数据,采用分类回归树（CART）和随机森林（RF）两种机器学习方法,结合地形因子、土壤类型、归一化植被指数,探究与宁夏南部地区土壤质地分布相关性较强的环境因素,并用两种机器学习预测该区土壤质地类型的空间分布,用剖面点验证集数据以及宁夏回族自治区海原县实测样点数据验证模型精度。【结果】（1）RF和CART对剖面点验证集土壤质地类型的预测正确率分别为 62.36%、55.29%,接收者操作特性（receiver operating characteristic,ROC）曲线下面积（area under roc curve,AUC）分别为0.7515、0.6933,对海原县122个实测样点的预测正确率分别为54.10%、48.36%,AUC分别为0.6599、0.5981,RF的预测精度高于CART。（2）该区土壤类型（ST）是与土壤质地空间分布相关性最强的环境因素,其次是高程（Ele）,高程越高,土壤质地越黏重。风力作用指数（WEI）和坡度（Slo）对土壤质地的影响较小。（3）研究区土壤质地类型以轻壤土为主,空间分布格局基本呈现为南部土壤质地黏重,北部土壤质地较轻。【结论】RF更适合预测宁夏南部地区土壤质地的空间分布,且充分利用历史数据,结合新的野外采样,可以达到预测制图的精度要求;土壤类型（ST）和高程（Ele）是与土壤质地空间分布相关性较强的环境因素。

Research on Spatial Distribution of Soil Texture in Southern Ningxia Based on Machine Learning

【Objective】Based on historical soil data, this paper studied the spatial variability of soil texture and its relationship with environmental factors in southern Ningxia by using machine learning.【Method】Classification and regression tree (CART), random forest (RF) and traditional statistical methods were used to explore the main environmental factors that affected the soil texture types and predict the spatial distribution of soil texture types in southern Ningxia, based on 428 soil profiles from the second soil survey in the 1980s, combined with topographic factors, soil types, and normalized vegetation index. The accuracy of the models were verified by the validating set of soil profiles and the soil samples in Haiyuan County, Ningxia.【Result】(1)The accuracy rates of RF and CART on the soil texture type of the verification set of soil profiles were 62.36% and 55.29%, respectively; the area under the receiver operating characteristic (ROC) curve of them (area under roc curve, AUC) were 0.7515 and 0.6933, respectively; the accuracy rates of them on soil samples in Haiyuan County were 54.10% and 48.36%, respectively; the AUC of them were 0.6599 and 0.5981 respectively. (2) Soil type (ST) was the most important predictor variable, followed by elevation (Ele). The higher elevation was, the heavier the soil texture was. The effects of wind exposition index (WEI) and slope (Slo) on soil texture were lower. (3)The results predicted by two methods showed a spatial distribution trend that the soil texture was heavy in the southern area but light in the northern area of southern Ningxia.【Conclusion】The prediction accuracy of RF for soil texture type in southern Ningxia was higher than CART. Making full use of historical data, combined with field sampling, could meet the accuracy requirements of digital mapping. In the loess region, soil types and elevation were the environmental factors which had strong correlation with spatial variation of soil texture.