基于BP神经网络的多年平均降雨量插值方法

Multi-Year Mean Precipitation In lation Method Based on on Neural Network Model

以广东省186个降雨站点多年平均降雨量为基础数据，在分析了多年平均降雨量的空间分布特征及其与经度、纬度、海拔高度、坡度、坡向的内在关系后，提出了一种基于DEM、GIS技术，利用BP神经网络模型进行插值的新方法（BPNNSI）。用没有参与建模的36个验证站点进行验证，结果表明；①BPNNSI方法的最大相对误差为-10．2％，平均相对误差为3．79％，插值结果与观测值的相关系数达到0．93，取得了较好的模拟效果。②从插值精度验证的5个指标（MRE、MMRE、MAE、C、误差分布范围）来看，该方法由于综合考虑降雨量的多种影响因素，因而都明显地优于IDW、KRIGING方法。因而，BPNNSI不仅能够用于降雨量的空间插值，而且还可以用于生成高精度的分布图，客观细致地反映降雨随其影响因素梯度变化的地带性特征。

By using precipitation dataset of 186 rain gauges in Guangdong Province, this paper analyzes the spatial distribution of multiyear mean precipitation and the relationships between rainfall and longitude, latitude, elevation, slope and direction of slope. This paper also provides a new method for multi-year mean precipitation interpolation based on DEM, GIS, and the back propagation neural network model （BPNNSI）, and the model is validated using data of 36 rain gauges, which are not used in the model calibration. The results of the model application in Guangdong Province reveal that the model can simulate the rainfall with high efficiency , （1）TheThe highest value of relative difference is -10. 2 %, the lowest value is 0. 06 % and the average value is 3. 79%; Correlation coefficient between the observed values and modeling values is 0. 93; （2）Compared with other models generally used （such as IDW and KRIGING）, the modeling and prediction precision of the model is the highest among them because of its complex capabilities of nonlinear mapping. So the method not only can be applied to surface interpolation of multi--year mean precipitation, but also can be used for generating high spatial resolution data sets, which can better reflect regional and spatial features of precipitation and explain the quantitative relation between precipitation and its effecting factors.