贝叶斯模型在土壤转换函数中的应用与适应性评价

Application and Adaptability Evaluation of Bayesian Model in Soil Transfer Functions

为了研究大型灌区节水改造后的区域农田生态环境效应中分布式水文模型空间参数的确定问题,通过内蒙古河套灌区解放闸灌域22个土壤水盐监测点110个土壤样本的采样与分析,利用贝叶斯神经网络(BNN)模型建立了河套灌区区域分层土壤特征参数与土壤水分特征曲线模型参数、特征含水率之间的土壤转换函数模型,并与已有的BP神经网络模型进行适应性比较及模型验证。结果表明,BP模型土壤转换函数的训练模拟精度优于BNN,但是在模拟预测方面,BNN模型普遍好于BP模型,而且模型输入因子数量对BP模型的精度影响较大,而BNN模型对于不同输入因子表现出很好的稳健性,BNN模型比传统的人工神经网络模型具有更好的适应性和预测效果,体现了土壤特征参数的空间随机性和结构性特征,而且预测的土壤水分特征曲线与实测和VG拟合结果更为接近,是一种具有广阔应用前景的区域土壤转换函数推求方法。

In order to s ecological influences of tudy the spatial parameters of the distributive hydrological models among the regional farmland under the condition of water-saving practices in large scale irrigation district, the Bayesian neural networks and back-propagation artificial neural network models were applied to establish regional pedotransfer function models. Based on the relationship of measured soil characteristic contents, soil particle percentage, organic matter and bulk density, the adaptability of these two kinds of ANN models were evaluated through simulated and predicted values statistically, accompanied with the SWRC figures. Results indicated that the BP and BNN were both feasible PTFs methods. The training simulated accuracy of traditional BP model was better than that of BNN. However, the predicted accuracy of BNN model generally was better than the BP model. Furthermore, the predictive accuracy of BP model was significantly influenced by the number of input factor groups. But there were little influences on different input factors of BNN model. So, the BNN showed good robustness for the simple inputs. Besides, the predicted SWRC was better fitted with measured and VG fitted curve than that of ANN. Thus, the BNN model was better than the traditional artificial neural network model. It had better adaptability in the pedotransfer function establishment when only soil particle distribution was used. All showing that the BNN method was a practical method for regional pedotransfer function establishment.