论文部分内容阅读
基于数字地面模型(Digital Terrain Model,DTM),同时考虑因子组合和分辨率构建土壤有机碳(SOC)最优估算模型。在7 100 km2范围内,选取了71个分辨率和22个地形因子中不多于5个因子的所有可能组合,构造了2 514 820个模型。采样点随机分为两组,6 362个训练样点构造数据挖掘模型,其他2 208个为验证样点。根据模型相关系数r值大小从中选取了不同个数因子组合以及相应分辨率的最优模型,并根据这些模型生成对应的土壤有机碳图。结果表明:单个地形因子模型和栅格大小之间的关系表现出多样化,并不是分辨率越高模型结果越好。单因子模型r值的大小并不能决定其在因子组合模型中的重要性。不同的因子及其组合有其特定的最适分辨率,最佳分辨率范围约为60~150 m。综合数据的存储空间和计算量、模型复杂度、预测精度以及空间表达能力,该地区最优模型由相对坡位、高程、归一化高程及多尺度山谷平坦指数等4个变量组成,对应分辨率为121.6 m。同时与多种克里格空间插值方法生成的土壤有机碳空间分布图进行了对比分析,发现无论几个变量的组合,其空间预测能力均较克里格空间插值方法更能表达SOC的空间变化,预测精度也较高。
Based on Digital Terrain Model (DTM), an optimal estimation model of soil organic carbon (SOC) was constructed considering both factor combination and resolution. Within the range of 7 100 km2, all possible combinations of not more than 5 factors out of 71 resolutions and 22 topographic factors were selected and 2 514 820 models were constructed. The sampling points were randomly divided into two groups, 6362 training samples were constructed data mining model, and the other 2 208 verification samples. According to the value of the model correlation coefficient r, the optimal combination of different number of factors and the corresponding resolution are selected, and the corresponding soil organic carbon map is generated according to these models. The results show that the relationship between the single terrain factor model and the grid size is diversified. The higher the resolution, the better the model result. The size of the one-factor model r does not determine its importance in the factor combination model. Different factors and their combinations have their specific optimum resolution, the best resolution range of about 60 ~ 150 m. Comprehensive data storage space and computational complexity, model complexity, prediction accuracy and spatial expression capabilities, the optimal model in the region by the relative slope, elevation, normalized elevation and multi-scale valley flat index and other four variables, corresponding to the resolution The rate is 121.6 m. At the same time, compared with the spatial distribution of soil organic carbon generated by a variety of Kriging interpolation methods, it was found that regardless of the combination of several variables, the spatial prediction ability of SOC is more than that of Kriging space interpolation method to express the spatial variation of SOC , The prediction accuracy is also higher.