Sahoo Goloka B, Ray Chittaranjan, Mehnert Edward, Keefer Donald A
Department of Civil and Environmental Engineering, University of Hawaii at Manoa, 2540, Dole Street, 383 Holmes Hall, Honolulu, HI 96822, USA.
Sci Total Environ. 2006 Aug 15;367(1):234-51. doi: 10.1016/j.scitotenv.2005.12.011. Epub 2006 Feb 7.
In this study, a feed-forward back-propagation neural network (BPNN) was developed and applied to predict pesticide concentrations in groundwater monitoring wells. Pesticide concentration data are challenging to analyze because they tend to be highly censored. Input data to the neural network included the categorical indices of depth to aquifer material, pesticide leaching class, aquifer sensitivity to pesticide contamination, time (month) of sample collection, well depth, depth to water from land surface, and additional travel distance in the saturated zone (i.e., distance from land surface to midpoint of well screen). The output of the neural network was the total pesticide concentration detected in the well. The model prediction results produced good agreements with observed data in terms of correlation coefficient (R=0.87) and pesticide detection efficiency (E=89%), as well as good match between the observed and predicted "class" groups. The relative importance of input parameters to pesticide occurrence in groundwater was examined in terms of R, E, mean error (ME), root mean square error (RMSE), and pesticide occurrence "class" groups by eliminating some key input parameters to the model. Well depth and time of sample collection were the most sensitive input parameters for predicting the pesticide contamination potential of a well. This infers that wells tapping shallow aquifers are more vulnerable to pesticide contamination than those wells tapping deeper aquifers. Pesticide occurrences during post-application months (June through October) were found to be 2.5 to 3 times higher than pesticide occurrences during other months (November through April). The BPNN was used to rank the input parameters with highest potential to contaminate groundwater, including two original and five ancillary parameters. The two original parameters are depth to aquifer material and pesticide leaching class. When these two parameters were the only input parameters for the BPNN, they were not able to predict contamination potential. However, when they were used with other parameters, the predictive performance efficiency of the BPNN in terms of R, E, ME, RMSE, and pesticide occurrence "class" groups increased. Ancillary data include data collected during the study such as well depth and time of sample collection. The BPNN indicated that the ancillary data had more predictive power than the original data. The BPNN results will help researchers identify parameters to improve maps of aquifer sensitivity to pesticide contamination.
在本研究中，开发了一种前馈反向传播神经网络（BPNN）并将其应用于预测地下水监测井中的农药浓度。农药浓度数据难以分析，因为它们往往受到高度审查。神经网络的输入数据包括含水层物质深度的分类指标、农药淋溶类别、含水层对农药污染的敏感性、样品采集时间（月）、井深、地面到水面的深度以及饱和带中的额外运移距离（即从地面到井筛中点的距离）。神经网络的输出是井中检测到的总农药浓度。模型预测结果在相关系数（R = 0.87）和农药检测效率（E = 89%）方面与观测数据达成了良好的一致性，并且在观测和预测的“类别”组之间也有很好的匹配。通过消除模型的一些关键输入参数，从R、E、平均误差（ME）、均方根误差（RMSE）以及农药出现“类别”组等方面考察了输入参数对地下水中农药出现的相对重要性。井深和样品采集时间是预测井的农药污染潜力最敏感的输入参数。这意味着开采浅层含水层的井比开采深层含水层的井更容易受到农药污染。发现施药后月份（6月至10月）的农药出现频率比其他月份（11月至4月）高2.5至3倍。BPNN用于对最有可能污染地下水的输入参数进行排序，包括两个原始参数和五个辅助参数。两个原始参数是含水层物质深度和农药淋溶类别。当这两个参数是BPNN的唯一输入参数时，它们无法预测污染潜力。然而，当它们与其他参数一起使用时，BPNN在R、E、ME、RMSE以及农药出现“类别”组方面的预测性能效率提高。辅助数据包括研究期间收集的数据，如井深和样品采集时间。BPNN表明辅助数据比原始数据具有更强的预测能力。BPNN的结果将有助于研究人员确定参数，以改进含水层对农药污染敏感性的地图。
