The soil in the sugarcane growing area of far north Queensland is often sodic (exchangeable sodium percentage—ESP > 6%). Gypsum therefore needs to be applied to reduce potential for land degradation. To accurately map ESP, a digital soil map (DSM) approach can be used. In this paper, we compare and contrast various aspects of DSM for mapping topsoil (0–0.3 m) ESP, including a suitable model (i.e. linear mixed model (LMM), Cubist and regression kriging (Cubist-RK)), usefulness of digital data (in combination or alone) and how many calibration data (i.e. n = 20, 30,…120) are required. We compare these with ordinary kriging (OK) of soil data and using prediction agreement (i.e. Lin's concordance—Lin's) and accuracy (root mean squared error—RMSE). We compare all of these results with a DSM derived from numerical clustering (fuzzy K-mean—FKM) of digital data to identify management zones (k = 2, 3, 4 and 5) and a conventional Soil Order map (k = 5 Orders). We do this by calculating mean squared prediction error (MSPE). Prediction of topsoil ESP by OK using 120 samples gave moderate agreement (Lin's = 0.72) with accuracy satisfactory given RMSE (3.69) was less than half standard deviation of measured ESP (½SD = 3.75). Moreover, a minimum number of 100 samples would be required for OK. However, when digital data were used to value add to soil data in models, the results were equivocal, given Cubist (Lin's = 0.74) and Cubist-RK (0.79) outperformed OK, while LMM (0.65) was inferior to OK. In addition, a smaller sample size (i.e. 70 and 60, respectively) was enough for Cubist and Cubist-RK to permit the development of accurate predictions of ESP given the RMSE was less than ½SD of measured ESP. Prediction of ESP (considering 120 samples) using only the γ-ray data (Lin's = 0.77) was superior to EC_a (0.72), however, using both in combination was best (0.79). The MSPE (n = 120) indicated creating DSM from clustering of digital data was best for k = 4 zones (MSPE = 27.60); however, Cubist-RK (13.40), Cubist (14.75), OK (15.56) and LMM (15.76) were able to provide better prediction of ESP. Nevertheless, all DSM generated smaller MSPE than a conventional Soil Order map (32.33). We recommend using Cubist-RK and both digital data, is the optimal approach to develop a DSM for application of gypsum to enable implementation of Six-Easy-Steps soil management guidelines for Proserpine.

中文翻译：

---

数字土壤图与常规土壤图在表层土壤可交换钠百分比管理中的比较

昆士兰州北部甘蔗种植区的土壤通常是钠质的（可交换钠含量——ESP > 6%）。因此，需要使用石膏来减少土地退化的可能性。为了准确地绘制 ESP，可以使用数字土壤图 (DSM) 方法。在本文中，我们比较和对比了 DSM 绘制表土 (0-0.3 m) ESP 的各个方面，包括合适的模型（即线性混合模型 (LMM)、立体主义和回归克里金法 (Cubist-RK)）、数字化的有用性数据（组合或单独）和多少校准数据（即n = 20, 30,…120) 是必需的。我们将这些与土壤数据的普通克里金法 (OK) 进行比较，并使用预测一致性（即 Lin 的一致性 - Lin's）和准确度（均方根误差 - RMSE）。我们将所有这些结果与从数字数据的数值聚类（模糊 K-mean-FKM）中得出的 DSM 进行比较，以识别管理区域（k  = 2、3、4 和 5）和传统的土壤顺序图（k  = 5 阶） ）。我们通过计算均方预测误差 (MSPE) 来做到这一点。使用 120 个样本通过 OK 预测表土 ESP 给出中等一致性 (Lin's = 0.72)，准确度令人满意，因为 RMSE (3.69) 小于测量 ESP 的一半标准偏差 (½ SD = 3.75)。此外，OK 至少需要 100 个样本。然而，当数字数据被用于模型中的土壤数据增值时，结果是模棱两可的，因为 Cubist (Lin's = 0.74) 和 Cubist-RK (0.79) 优于 OK，而 LMM (0.65) 不如 OK。此外，较小的样本量（即分别为 70 和 60）足以让 Cubist 和 Cubist-RK 允许开发准确的 ESP 预测，因为 RMSE 小于测量的 ESP 的 ½ SD。仅使用 γ 射线数据 (Lin's = 0.77) 预测 ESP（考虑 120 个样本）优于 EC _a (0.72)，但是，两者结合使用效果最好 (0.79)。MSPE ( n  = 120) 表明从数字数据的聚类中创建 DSM 最适合k = 4 个区域（MSPE = 27.60）；然而，Cubist-RK (13.40)、Cubist (14.75)、OK (15.56) 和 LMM (15.76) 能够提供更好的 ESP 预测。尽管如此，所有 DSM 生成的 MSPE 都比传统的土壤顺序图 (32.33) 更小。我们建议使用 Cubist-RK 和这两种数字数据，这是开发用于石膏应用的 DSM 以实现 Proserpine 的六步简单土壤管理指南的最佳方法。