Abstract Prediction of soil moisture is critical for water resources management. With Global Precipitation Measurement (GPM) and Soil Moisture Active Passive (SMAP) satellites by NASA, spatio-temporal interrelation between rainfall and soil moisture fields (at different extents) will be of great value for satellite product calibration/validation and other hydrologic science investigations. In this study, we explored a non-parametric evolutionary algorithm for prediction of soil moisture from a time series of spatially-distributed rainfall across multiple weather locations under two different hydro-climatic regions. A new genetic algorithm-based hidden Markov model (HMMGA) was developed to estimate long-term soil moisture dynamics at different soil depths using precipitation data as a proxy. Also, we tested transposability of our approach across time under different climatic conditions. To test the new approach, we selected two different soil moisture fields, Oklahoma (130 km × 130 km) and Illinois (300 km × 500 km), during 1995–2009 and 1994–2010, respectively. We found that the newly developed framework performed well in predicting soil moisture dynamics at different spatial extents. Although our approach has limitations in predicting daily values, it estimates well the weekly soil moisture across the spatial and temporal domains with predictable uncertainties. Furthermore, this approach could provide advantages for good transposability under different weather conditions compared to those of physically-based hydrological models. Overall, our suggested approach could predict weekly soil moisture estimates with precipitation and soil moisture histories and showed the potential of transposability under different weather and land surface conditions. Since the proposed algorithm requires only precipitation (and historical soil moisture data) from existing, established weather stations, it can serve an attractive alternative that can forecast soil moisture using climate change scenarios.

中文翻译：

---

用于预测土壤水分动力学的非参数进化算法的发展

摘要 土壤水分预测对于水资源管理至关重要。借助 NASA 的全球降水测量 (GPM) 和土壤水分主动被动 (SMAP) 卫星，降雨量和土壤水分场（不同程度）之间的时空相互关系将对卫星产品校准/验证和其他水文科学研究具有重要价值. 在这项研究中，我们探索了一种非参数进化算法，用于根据两个不同水文气候区下多个天气位置的空间分布降雨时间序列预测土壤水分。开发了一种新的基于遗传算法的隐马尔可夫模型 (HMMGA)，以使用降水数据作为代理来估计不同土壤深度的长期土壤水分动态。还，我们在不同气候条件下测试了我们的方法随时间的可转换性。为了测试新方法，我们分别在 1995-2009 年和 1994-2010 年期间选择了两个不同的土壤湿度场，俄克拉荷马州（130 公里 × 130 公里）和伊利诺伊州（300 公里 × 500 公里）。我们发现新开发的框架在预测不同空间范围的土壤水分动态方面表现良好。尽管我们的方法在预测每日值方面存在局限性，但它可以很好地估计时空域上的每周土壤水分，具有可预测的不确定性。此外，与基于物理的水文模型相比，这种方法可以为不同天气条件下的良好转座提供优势。全面的，我们建议的方法可以通过降水和土壤水分历史预测每周土壤水分估计值，并显示在不同天气和地表条件下的转座潜力。由于所提出的算法只需要来自现有已建立气象站的降水（和历史土壤湿度数据），因此它可以作为一种有吸引力的替代方案，可以使用气候变化情景预测土壤湿度。