Soil water repellency (SWR) is a natural phenomenon occurring in soils throughout the world, which impacts upon ecosystem services at multiple temporal and spatial scales (nano to ecosystem scale). In pastures, the development of SWR is primarily determined by the cycling of hydrophobic materials at the soil surface, and is controlled by climate, soil and water management, and soil properties. The complex interactions between these factors make it an intricate system to understand and model. Detailed spatiotemporal characterization of the surface moisture and biomass in pastoral ecosystems would allow for a better understanding of this phenomenon. Normalized Difference Vegetation Index (NDVI) and Synthetic Aperture Radar (SAR) backscatter are good predictors for surface biomass and soil moisture, respectively. Machine learning on remote sensing time series (TS) data shows promise to predict the occurrence of SWR in pastures. This study evaluates the ability of remote sensing TS data to predict the occurrence of SWR in New Zealand pastures, using three machine learning algorithms. Soil water repellency data were collected from 58 pastoral sites. Machine learning models were trained and cross-validated on a monthly aggregated remote sensing and water deficit TS data to predict SWR level. Prediction output from artificial neural networks (ANN), random forest (RF), and support vector machine (SVM) were compared using root mean squared error (RMSE). When using NDVI TS data from 58 site as predictors of SWR, SVM and RF (RMSE = 0.82 and 0.87, respectively) outperformed ANN (RMSE = 1.23). Random forest was used to map SWR magnitude over Hawke’s Bay region in the North Island of New Zealand, and the overall accuracy was equal to 86%. This study is the first investigation implicating remote sensing TS data to predict the occurrence of SWR at the regional scale. Mapping the potential SWR will aid in identifying critical zones of SWR, to attenuate its effect on pastures through adapted management.

土壤憎水性（SWR）是一种在世界各地土壤中发生的自然现象，在多个时空尺度（纳米到生态系统尺度）上都会影响生态系统服务。在牧场中，SWR的发展主要取决于土壤表面疏水性物质的循环，并受气候，土壤和水的管理以及土壤性质的控制。这些因素之间的复杂相互作用使它成为理解和建模的复杂系统。牧区生态系统中表面水分和生物量的详细时空特征可以更好地了解这一现象。归一化植被指数（NDVI）和合成孔径雷达（SAR）反向散射分别是表面生物量和土壤水分的良好预测指标。对遥感时间序列（TS）数据的机器学习表明，有望预测牧场中SWR的发生。这项研究使用三种机器学习算法评估了遥感TS数据预测新西兰牧场SWR发生的能力。从58个牧区收集了土壤憎水数据。对机器学习模型进行了训练，并根据每月汇总的遥感数据和水分亏缺TS数据进行了交叉验证，以预测SWR水平。使用均方根误差（RMSE）比较了人工神经网络（ANN），随机森林（RF）和支持向量机（SVM）的预测输出。当使用58个站点的NDVI TS数据作为SWR，SVM和RF（分别为RMSE = 0.82和0.87）的预测指标时，其性能优于ANN（RMSE = 1.23）。随机森林被用来绘制新西兰北岛霍克湾地区的SWR幅值，总体准确度等于86％。这项研究是第一个涉及遥感TS数据以预测区域尺度SWR发生的调查。绘制潜在的SWR有助于确定SWR的关键区域，从而通过适当的管理来减轻其对牧场的影响。