西北旱区幼龄果园土壤蒸发特征与模拟方法

Soil Evaporation Characteristics and Multi-model Simulation Methods in Young Orchards in Arid Region of Northwest China

摘要:

为明确西北旱区幼龄果园土壤蒸发特征及不同模拟方法的适用性,本文以甘肃武威石羊河流域幼龄李树和杏树果园为研究对象,基于土壤蒸发量、土壤含水率、土壤温度、土壤水势、气象因素和叶面积指数等多源观测资料,分析了幼龄果园土壤蒸发变化规律,并比较了FAO-56土壤蒸发系数、Ritchie、随机森林(RF)、极端梯度提升机(XGBoost)和轻量梯度提升机(LightGBM)模型模拟效果。结果表明,李树和杏树果园土壤蒸发量分别为0.19~7.41 mm/d和0.06~7.00 mm/d,均值分别为2.00 mm/d和2.03 mm/d,均表现出明显的日尺度波动。土壤蒸发主要受参考作物蒸散量、太阳辐射、风速及浅层土壤含水率呈不同程度相关影响,其中李树果园与大气蒸发需求和辐射条件的同步性较强,杏树果园与浅层土壤水势和表层含水率的关系更密切。不同模型均能在一定程度上反映土壤蒸发平均水平,但对日尺度波动和高蒸发样本的刻画能力差异明显。传统模型模拟范围偏窄,对高蒸发量样本存在低估,其中FAO-56土壤蒸发系数模型在李树果园中模拟值仅为0.53~4.60 mm/d,RMSE为0.72 mm/d。相比之下,XGBoost模型对土壤蒸发变化幅度和高值样本的表征能力最强,在李树果园中模拟值为0.24~7.35 mm/d,R2、NSE、MAE和RMSE分别为0.95、0.94、0.25 mm/d和0.35 mm/d,较FAO-56方法RMSE降低约51.4%;在杏树果园中模拟值为0.11~6.97 mm/d,R2、NSE、MAE和RMSE分别为0.95、0.93、0.23 mm/d和0.36 mm/d。LightGBM和RF模型精度次之,但整体仍优于FAO-56土壤蒸发系数模型和Ritchie模型。综上,基于多源气象和土壤水热信息的机器学习模型可提高幼龄果园土壤蒸发日尺度模拟精度,其中XGBoost综合表现最优,可为旱区幼龄果园控蒸保墒和精准灌溉提供理论支持。

Abstract:

Aiming to clarify soil evaporation dynamics and the applicability of different simulation methods in young orchards in arid Northwest China, young plum and apricot orchards in the Shiyang River Basin, Wuwei, Gansu Province, were selected as study sites. Based on multi-source observations, including soil evaporation, soil water content, soil temperature, soil water potential, meteorological factors and leaf area index, the variation characteristics of soil evaporation in young orchards were analyzed. The performance of the simplified FAO-56 soil evaporation coefficient method, the Ritchie model, random forest (RF), extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM) was then evaluated. The results showed that soil evaporation ranged from 0.19 mm/d to 7.41 mm/d in the plum orchard and from 0.06 mm/d to 7.00 mm/d in the apricot orchard, with mean values of 2.00 mm/d and 2.03 mm/d, respectively, indicating pronounced daily fluctuations. Soil evaporation was jointly affected by reference crop evapotranspiration, solar radiation, wind speed, and shallow soil water conditions. Soil evaporation in the plum orchard was more closely synchronized with atmospheric evaporative demand and radiation conditions, whereas that in the apricot orchard was more closely associated with shallow soil water potential and surface soil water content. All models reproduced the mean soil evaporation level to some extent, but their ability to capture daily fluctuations and high-evaporation samples differed markedly. The traditional models had relatively narrow simulation ranges and tended to underestimate high soil evaporation values. For example, the simulation range of the simplified FAO-56 soil evaporation coefficient method in the plum orchard was only 0.53~4.60 mm/d, with an RMSE of 0.72 mm/d. In contrast, the XGBoost model showed the strongest ability to represent the variation range and high-value samples of soil evaporation. In the plum orchard, the simulation range of XGBoost was 0.24~7.35 mm/d, and its R2, NSE, MAE, and RMSE were 0.95, 0.94, 0.25 mm/d, and 0.35 mm/d, respectively, with the RMSE reduced by approximately 51.4% compared with the FAO-56 method. In the apricot orchard, the simulation range of XGBoost was 0.11~6.97 mm/d, and its R2, NSE, MAE, and RMSE were 0.95, 0.93, 0.23 mm/d, and 0.36 mm/d, respectively. LightGBM and RF showed intermediate accuracy, but both outperformed the FAO-56 soil evaporation coefficient method and the Ritchie model. In conclusion, machine learning models integrating multi-source meteorological and soil hydrothermal information can improve the accuracy of daily soil evaporation estimation in young orchards, with XGBoost showing the best overall performance. These results can provide a methodological reference for soil evaporation estimation, evaporation control, soil moisture conservation, and precision irrigation management in young orchards in arid regions.