利用BP神经网络由特征气象要素预测土壤湿度

Applying BP Neural Network to Estimate Soil Moisture with Meteorologic Parameters

土壤湿度的实测数据量因地面观测网的稀疏不均和观测时间的不连续而十分有限,而其作为气象、农业、水文、环境等学科领域的重要研究内容,数据量的匮乏直接影响了研究工作的顺利进行。地面常规气象数据的观测频率较高(逐日/时),提供了丰富的大气及土壤状态信息,从地气交互作用的普遍性出发,对气象要素与土壤湿度之间的作用关系进行研究,拟借助人工神经网络良好的函数模拟能力,建立以多气象要素为网络输入、以土壤湿度为网络输出的BP神经网络。通过主成分分析筛选特征要素、选择训练函数、确定合理的隐层神经元个数等来精细化网络。以甘肃省2008年8、9月份的AB报(土壤湿度数据)和A报(气象观测数据)资料进行了实验,建立BP神经网络,最终获得了较好的土壤湿度预测结果。

The data of soil moisture is quite limited because of the sparse observation network and discontinuous observation time.As one important element of meteorology,agriculture,hydrology,and environment,the soil moisture data＇s insufficiency has a directly pernicious influence on the related research work.On the other hand,the routine meteorologic data has a high observation frequency（per day/hour） and provides rich information on atmosphere and soil at the same time.In view of the universal interaction between land and atmosphere,we try to do some research in the relationship between soil moisture and meteorologic parameters.Through artificial neural network＇s excellent capability in function simulation,we manage to establish a BP neural network with multi-meteorologic parameters being as the input and soil moisture as the output.The principle component analysis is used to choose representative meteorologic parameters,training algorithm and define a right number for hidden layer neurons.With AB report（data for soil moisture） and A report（meteorologic observations） of Gansu province in August and September,2008,China,BP neutral network is built and good retrieval results are achieved.