基于平均影响值算法和BP神经网络的根区土壤湿度估算

为了探讨基于地表特征信息应用人工神经网络计算根区土壤湿度(root zone soil moisture, RZSM)的可行性，利用中国境内4个典型区域的农田生态系统野外台站的地表和根区土壤水分实测数据，结合6种气象数据和2种植被指数数据构建了不同深度根区土壤湿度的反向传播(back propagation, BP)神经网络计算模型，采用决定系数R²、均方根误差(root mean squared error, RMSE)和平均绝对误差(mean absolute error, MAE)3个指标来评估4个站点不同土层深度的模型性能，并使用平均影响值(mean impact value, MIV)算法得到9个地表特征变量的重要性。结果表明：在20～90 cm和20～100 cm深度，模型的平均R²值分别为0.79、0.69、0.66、0.56、0.51和0.47;RMSE为1.91%、2.17%、2.51%、2.71%、2.82%和3.08%;MAE为1.44%、1.61%、1.75%、1.89%、2.04%和2.35%,表明BP神经网络模型能...

Calculation of Root Zone Soil Moisture Using MIV-BP Neural Networks

In order to explore the feasibility of using the artificial neural network method to calculate the RZSM based on the surface information, the BP neural network models were constructed to estimate the RZSM at different soil depths using the measured soil moisture data, meteorological data and vegetation index data at four typical sites in China. The coefficient of determination (R2), the root mean square error (RMSE) and the mean absolute error (MAE) were used to evaluate the performance of the models at different soil depths. The mean impact value (MIV) was used to obtain the importance of the input variables. The results show that at depths of 20~90/100cm, the average R2 values of the model are 0.79, 0.69, 0.66, 0.56, 0.51 and 0.47, respectively; RMSE is 1.91%, 2.17%, 2.51%, 2.71%, 2.82% and 3.08%; MAE is 1.44%, 1.61% 1.75%, 1.89%, 2.04%, and 2.35%, which indicated that the BP Neural Networks method was able to fit the RZSM variability at different stations with different climates and soil type, but the performance and accuracy of RZSM decrease with the increase of soil depth. The results also indicated the model capability of RZSM calculation decreased as the increase of soil depth. The MIV results showed that SSM is the most important feature in the model and the interaction between multiple ground features is an auxiliary condition, also, the effects of different features on RZSM in different climate zones and soil types are inconsistent.