Abstract Soil maps are an important tool for agricultural planning and land management. Digital techniques have been used to create soil maps. However, most studies did not explore drainage network (DN) information on prediction models, which are related to soil variability. Thus, this study aims to evaluate the contribution of DN to predict soil classes using digital soil mapping techniques. We used a conventional soil class map (1:20,000) and environmental variables, such as drainage and relief attributes and satellite images, aiming to extrapolate the soil map to a larger area. The work was conducted in Sao Paulo State, Brazil. We created a point grid with 30 × 30 m resolution to extract the soil and variables information. We used these data to calibrate a random forest model along with cross-validation to optimize the model selection. The predicted soil classes for the 53,800-ha study area were determined on two levels according to the World Reference Base (WRB) soil classification system. The first level considered only soil groups (i.e. Acrisol and Ferralsol), while the second level considered the soil group and a qualifier (i.e. Chromic Acrisol and Rhodic Acrisol). We validated the maps using other conventional soils maps (internal validation) and field sampling points (external validation). After extrapolating the soil map, we validated the model s performance using field observations. In this case, the method reached an accuracy of 0.56 and kappa of 0.31 for the soil’s first level, and 0.38 and 0.25 for the second level. Regosols and Cambisols prediction was underestimated, lowering the accuracy and kappa results on the validation. However, Ferralsols reached accuracy and Acrisols reached around 70% accuracy. The drainage related attributes had the highest contribution to the model’s performance (accuracy = 56%) and improved the soil map extrapolation.

中文翻译：

---

基于专家的地图和高度详细的地表排水模型，以支持数字土壤制图

摘要 土壤图是农业规划和土地管理的重要工具。数字技术已被用于创建土壤图。然而，大多数研究并未探索与土壤变异性相关的预测模型的排水网络 (DN) 信息。因此，本研究旨在评估 DN 对使用数字土壤绘图技术预测土壤类别的贡献。我们使用了传统的土壤类图 (1:20,000) 和环境变量，例如排水和地形属性以及卫星图像，旨在将土壤图外推到更大的区域。这项工作是在巴西圣保罗州进行的。我们创建了一个 30 × 30 m 分辨率的点网格来提取土壤和变量信息。我们使用这些数据来校准随机森林模型以及交叉验证以优化模型选择。根据世界参考基地 (WRB) 土壤分类系统，53,800 公顷研究区的预测土壤类别在两个级别上确定。第一级只考虑土壤组（即 Acrisol 和 Ferralsol），而第二级考虑土壤组和限定词（即 Chromic Acrisol 和 Rhodic Acrisol）。我们使用其他传统土壤图（内部验证）和现场采样点（外部验证）验证了这些地图。在外推土壤图后，我们使用实地观察验证了模型的性能。在这种情况下，该方法对于土壤的第一级达到了 0.56 和 0.31 的准确度，对于第二级达到了 0.38 和 0.25。Regosols 和 Cambisols 预测被低估，降低了验证的准确性和 kappa 结果。然而，Ferralsols 达到了准确度，而 Acrisols 达到了 70% 左右的准确度。排水相关属性对模型性能的贡献最大（精度 = 56%），并改进了土壤图外推。