INFOSOL, INRA, 45075 Orleans, France; Department of Geoscience, Environment & Society, Université Libre de Bruxelles, Brussels, Belgium.
URSOLS, INRA, 45075 Orléans, France.
Sci Total Environ. 2016 Dec 15;573:1352-1369. doi: 10.1016/j.scitotenv.2016.07.066. Epub 2016 Jul 16.
This work presents the first GlobalSoilMap (GSM) products for France. We developed an automatic procedure for mapping the primary soil properties (clay, silt, sand, coarse elements, pH, soil organic carbon (SOC), cation exchange capacity (CEC) and soil depth). The procedure employed a data-mining technique and a straightforward method for estimating the 90% confidence intervals (CIs). The most accurate models were obtained for pH, sand and silt. Next, CEC, clay and SOC were found reasonably accurate predicted. Coarse elements and soil depth were the least accurate of all models. Overall, all models were considered robust; important indicators for this were 1) the small difference in model diagnostics between the calibration and cross-validation set, 2) the unbiased mean predictions, 3) the smaller spatial structure of the prediction residuals in comparison to the observations and 4) the similar performance compared to other developed GlobalSoilMap products. Nevertheless, the confidence intervals (CIs) were rather wide for all soil properties. The median predictions became less reliable with increasing depth, as indicated by the increase of CIs with depth. In addition, model accuracy and the corresponding CIs varied depending on the soil variable of interest, soil depth and geographic location. These findings indicated that the CIs are as informative as the model diagnostics. In conclusion, the presented method resulted in reasonably accurate predictions for the majority of the soil properties. End users can employ the products for different purposes, as was demonstrated with some practical examples. The mapping routine is flexible for cloud-computing and provides ample opportunity to be further developed when desired by its users. This allows regional and international GSM partners with fewer resources to develop their own products or, otherwise, to improve the current routine and work together towards a robust high-resolution digital soil map of the world.
这项工作呈现了法国的第一张全球土壤图谱 (GSM) 产品。我们开发了一种自动程序，用于绘制主要土壤特性（粘土、粉砂、砂、粗粒元素、pH 值、土壤有机碳 (SOC)、阳离子交换容量 (CEC) 和土壤深度）。该程序采用了数据挖掘技术和一种简单的方法来估计 90%置信区间 (CI)。最准确的模型是 pH 值、砂和粉砂。其次，CEC、粘土和 SOC 的预测结果也相当准确。粗粒元素和土壤深度是所有模型中最不准确的。总的来说，所有模型都被认为是稳健的；这方面的重要指标是 1）校准和交叉验证集之间模型诊断的差异很小，2）无偏的平均预测值，3）预测残差的空间结构与观测值相比更小，4）与其他已开发的全球土壤图谱产品相比性能相似。然而，所有土壤特性的置信区间 (CI) 都相当宽。随着深度的增加，中位数预测变得越来越不可靠，CI 随深度增加而增加。此外，模型准确性和相应的 CI 取决于感兴趣的土壤变量、土壤深度和地理位置。这些发现表明，CI 与模型诊断一样具有信息性。总之，所提出的方法对大多数土壤特性产生了相当准确的预测。最终用户可以根据不同的目的使用这些产品，正如一些实际示例所展示的那样。该制图程序具有灵活性，可以进行云计算，并为其用户提供进一步开发的充分机会。这使得资源较少的区域和国际 GSM 合作伙伴能够开发自己的产品，或者改进当前的常规程序，并共同努力构建一个稳健的、高分辨率的世界数字土壤图谱。
