Management of micronutrient levels in soils must be done carefully to avoid their deficiency or toxicity to plants. The laboratory determination of micronutrient contents is time-consuming, expensive and generates chemical wastes, making it difficult for soil surveys required in precision agriculture, especially in tropical countries. While proximal sensors like portable X-ray fluorescence (pXRF) spectrometry have been successfully used to predict contents of soil available macronutrient, little effort has focused on micronutrients, especially involving a large dataset, soils weathering degree and a practical application of the predictions. This study aimed to use pXRF data for the prediction of available micronutrients in 1514 samples from variable soil classes (from Entisols to Oxisols) from seven Brazilian states using machine learning algorithms and to assess the influence of soil weathering degree on such prediction models. The soil samples were collected from both surface (A) and subsurface (B or C) horizons of various soil classes under several land uses, and with varying parent materials. Available B, Cu, Fe, Mn, and Zn were predicted via stepwise multiple linear regression (SMLR), support vector machine (SVM), extreme gradient boosting (XGB), and random forest (RF) algorithms and subsequently validated. The best prediction models were classified according to micronutrient availability classes (categorical validation). Adequate predictions were achieved for Cu: R² = 0.80; RPD = 2.28; Mn: 0.68; 1.76; and Zn: 0.68; 1.70. Predictions of B, Cu, Fe, Mn, and Zn availability classes yielded overall accuracy of 0.90, 0.65, 0.67, 0.73, and 0.53, respectively. Summarily, pXRF data in conjunction with prediction models can be an effective and rapid method to determine available Cu, Mn, and Zn. Soil weathering degree must be considered on such predictions as they strongly influence model accuracy.

必须谨慎管理土壤中的微量营养素水平，以避免它们缺乏或对植物产生毒性。微量营养素含量的实验室测定耗时、昂贵且会产生化学废物，使得精准农业所需的土壤调查变得困难，尤其是在热带国家。虽然诸如便携式 X 射线荧光 (pXRF) 光谱法等近端传感器已成功用于预测土壤可用常量营养素的含量，但很少有人关注微量营养素，尤其是涉及大型数据集、土壤风化程度和预测的实际应用。本研究旨在使用 pXRF 数据使用机器学习算法预测来自巴西七个州的可变土壤类别（从 Entisols 到 Oxisols）的 1514 个样本中的可用微量营养素，并评估土壤风化程度对此类预测模型的影响。土壤样品是从不同土地用途下的各种土壤类别的地表 (A) 和地下 (B 或 C) 层收集的，并且具有不同的母体材料。可用的 B、Cu、Fe、Mn 和 Zn 通过逐步多元线性回归 (SMLR)、支持向量机 (SVM)、极端梯度提升 (XGB) 和随机森林 (RF) 算法进行预测并随后进行验证。最佳预测模型根据微量营养素可用性类别（分类验证）进行分类。对 Cu 进行了充分的预测：R 和不同的母材。可用的 B、Cu、Fe、Mn 和 Zn 通过逐步多元线性回归 (SMLR)、支持向量机 (SVM)、极端梯度提升 (XGB) 和随机森林 (RF) 算法进行预测并随后进行验证。最佳预测模型根据微量营养素可用性类别（分类验证）进行分类。对 Cu 进行了充分的预测：R 和不同的母材。可用的 B、Cu、Fe、Mn 和 Zn 通过逐步多元线性回归 (SMLR)、支持向量机 (SVM)、极端梯度提升 (XGB) 和随机森林 (RF) 算法进行预测并随后进行验证。最佳预测模型根据微量营养素可用性类别（分类验证）进行分类。对 Cu 进行了充分的预测：R²  = 0.80；RPD = 2.28；锰：0.68；1.76; 和锌：0.68；1.70。B、Cu、Fe、Mn 和 Zn 可用性等级的预测分别产生了 0.90、0.65、0.67、0.73 和 0.53 的总体准确度。总之，结合预测模型的 pXRF 数据可以成为确定可用 Cu、Mn 和 Zn 的有效且快速的方法。在进行此类预测时必须考虑土壤风化程度，因为它们强烈影响模型精度。