基于贝叶斯最大熵和多源数据的作物需水量空间预测

作物需水量是灌溉工程规划、设计和管理的重要基础数据,充分利用多源数据和先验知识,快速经济地获取精度较高的区域作物需水量对于区域水资源的优化配置具有重要意义。为精确预测作物需水量,该文以长系列实际监测和校核作物系数后计算得到的作物需水量为硬数据,利用硬数据确定获得最大熵的约束条件,根据软数据获取渠道的不同(部分年份缺失的站点数据、文献中获得的数据、利用灌溉试验数据库中的作物需水量资料,采用协同克立格方法获得的数据、考虑主要地形因子和主要气象要素的影响,采用主成分分析和地理加权回归(geographically weighted regression,GWR)方法获得作物需水量数据以及遥感数据),提出不同来源软数据的概率密度函数表达方法,采用贝叶斯最大熵(Bayesian maximum entropy,BME)方法对不同来源的作物需水量信息进行有机整合。结果表明:除硬数据+文献软数据外,其他数据整合呈现一致结果。华北地区冬小麦作物需水量在豫南地区较小,中部地区黄河北岸有连片的相对高值区,山东需水量相对较高,冀东北的乐亭、唐山附近有相对低值区。除硬数据+文献软数据比不整合的精度低9.41%外,其他软数据源均可不同程度地提高整合效果,硬数据+克立格软数据、硬数据+GWR软数据和硬数据+除文献数据外的其他软数据分别比不整合的精度提高85.33%、85.75%和91.69%。对考虑地形、气象等要素的多源数据进行整合可更好地反映冬小麦作物需水量空间分布的细节,显著提高估算精度,为稀疏监测站点地区水土资源的精准管理和优化配置提供数据支撑。

Spatial prediction of crop water requirement based on Bayesian maximum entropy and multi-source data

Abstract: Crop water requirement is an important basic data for planning, design and management of irrigation engineering. Obtaining high-precision regional crop water requirement prediction using multi-source data and the priori knowledge has great significance for optimal allocation of regional water resources. In the paper, multisource data was integrated using the Bayesian maximum entropy (BME) method for crop water requirement prediction. A long series of crop water requirement measured and calculated by using the crop coefficients adjusted for actual measurement, were taken as the hard data. The soft data included the missing data in partial years for some stations, literature data, the Kriging interpolation data considering the main influence elements of crop water requirement, the crop water requirement data based on the principal component analysis (PCA) and geographically weighted regression (GWR) method, and the remote sensing data. For the soft data from different sources, the expression method of probability density function was put forward and the crop water requirement information from different sources was well integrated using the BME method. Hard data for the period of 1954-2013were collected from measurements from the irrigation stations in North China. Soft data for winter wheat in North China were also collected by searching literatures and the others. The results showed that spatial distribution of crop water requirement in North China was almost consistent for the hard data, hard data + Kriging soft data, hard data + GWR soft data and hard data + the soft data except for literature data. In the southern Henan had smaller crop water requirement, but the middle part (northern part of the Yellow River) of the North China was relatively high. The crop water requirement was relatively high in Shandong but low in the northeast of Hebei such as Leting, Tangshan. The results from hard data + literature soft data were slightly different from the others and the difference might be because the time periods used were different. In general, the integration accuracy of hard data + literature soft data was 9.41% lower than that based on hard data only. Hard data integrated with the other soft data could improve the integration effect. In particular, the integration accuracy of hard data + Kriging soft data, hard data + GWR soft data and hard data + the soft data except for literature data increased by 85.33%, 85.75% and 91.69%, respectively. The integration of multi-source data through considering the terrain, meteorological factors and etc, can could better reflect the spatial distribution of crop water requirements for winter wheat, and significantly improve the estimation accuracy of crop water requirement for winter wheat. The presented method provided the most important basic data for the precise management and optimization of water and soil resources in the region with sparse monitoring stations. In the paper, we need pay attention to some questions in the soft data processing. The partially missing data of some stations need amend the variance calculation results. In order to avoid the agglomeration phenomenon, the selection of interpolation data need adopt the method of random sampling, and at the same time, the distance between adjacent samplings must be limited, should not be too far or near, and 20 km was advisable. In order to avoid too big error and uncertainty, the literature data must be screened and pretreatment, otherwise, the integration effect may be affected.