Abstract Recently, portable X-ray fluorescence (pXRF) spectrometer and visible near-infrared (Vis-NIR) spectroscopy are increasingly being applied for soil types and attributes prediction, but a few works have used them combined in tropical regions. Thus, this work aimed at analyzing models’ performance when predicting soil types at subgroup taxonomic level via pXRF and Vis-NIR separately and together. 315 soil samples were collected in both A and B horizons in three important Brazilian states. Samples undergone laboratorial analyses for soil classification and were submitted to pXRF and Vis-NIR (350–2500 nm) analyses. Vis-NIR spectral data preprocessing was evaluated utilizing Savitzky-Golay (WT) and Savitzky-Golay with Binning (WB) methods. Four classification algorithms were employed in modeling: Support Vector Machine with Linear (SVM-L) and Radial (SVM-R) kernel, C5.0, and Random Forest (RF). Predictions were made using only B horizon and using A + B horizon data. Overall accuracy and Cohen’s Kappa index evaluated model quality. Both sensors displayed efficacy in soil types prediction. A + B horizons data combined using pXRF + Vis-NIR via SVM-R (WT and WB) delivered accurate predictions (89.32% overall accuracy and 0.75 Kappa index), but the best predictions were achieved using only B horizon data via pXRF with RF, pXRF + Vis-NIR (WT) with RF, pXRF + Vis-NIR (WB) with C5.0, and pXRF + Vis-NIR (WB) with RF (89.23% overall accuracy and 0.80 Kappa index). For tropical soils, soil subgroup prediction using only B horizon data obtained by pXRF in tandem with RF algorithm may be a viable alternative to assist in soil classification, especially when the acquisition of Vis-NIR is not possible.

中文翻译：

---

通过便携式 X 射线荧光和可见光近红外光谱预测土壤亚群

摘要 近年来，便携式X射线荧光（pXRF）光谱仪和可见光近红外（Vis-NIR）光谱仪越来越多地应用于土壤类型和属性预测，但在热带地区也有一些工作将它们结合起来。因此，这项工作旨在分析模型在通过 pXRF 和 Vis-NIR 分别和一起在亚组分类水平上预测土壤类型时的性能。在巴西三个重要州的 A 层和 B 层收集了 315 个土壤样品。样品经过实验室分析以进行土壤分类，并提交给 pXRF 和 Vis-NIR（350-2500 nm）分析。使用 Savitzky-Golay (WT) 和 Savitzky-Golay with Binning (WB) 方法评估了 Vis-NIR 光谱数据预处理。建模中采用了四种分类算法：具有线性 (SVM-L) 和径向 (SVM-R) 内核、C5.0 和随机森林 (RF) 的支持向量机。仅使用 B 地平线和 A + B 地平线数据进行预测。总体准确率和 Cohen's Kappa 指数评估模型质量。两种传感器都显示出土壤类型预测的有效性。通过 SVM-R（WT 和 WB）使用 pXRF + Vis-NIR 组合的 A + B 视界数据提供了准确的预测（89.32% 的总体准确度和 0.75 Kappa 指数），但最好的预测是通过 pXRF 和 RF 仅使用 B 视界数据实现的、pXRF + Vis-NIR (WT) 与 RF、pXRF + Vis-NIR (WB) 与 C5.0，以及 pXRF + Vis-NIR (WB) 与 RF（89.23% 的整体准确度和 0.80 Kappa 指数）。对于热带土壤，仅使用 pXRF 获得的 B 层位数据与 RF 算法结合进行土壤亚群预测可能是辅助土壤分类的可行替代方法，