Epsilon Archive for Student Projects

Abstract

The aim of this thesis was to explain variations in pesticide leaching from the pesticide properties (DT50, koc, log Pow, Sw and combinations of these) using multiple linear regression and artificial neural networks. The data came mainly from Vemmenhög, a catchment nine square kilometres in size dominated by agriculture, located in the south of Sweden. The analysed period is May to November 1997-2003. The artificial neural network, a feed-forward back propagation network, did not work in this case. For the regression analysis, a stepwise selection was used. Analyses were performed both on data where all zero-losses were excluded and on data where substances used in low amounts were excluded. Excluding the pesticides that were applied in low amounts gave better results than excluding those with zero-loss. With loss rate as a response variable, it was possible to find significant functions explaining up to 99% of the variability for individual years. The combination of variables in the functions with the highest degree of explanation (r2) differed for different years, but DT50/koc was the most frequently occurring variable. Grouping the years, the best significant function for 1997-2003 (excluding pesticides used in low amounts) contained DT50/koc and log Pow, with an r2 value of 70% (P<0.0001). It was generally not possible to use the formulas to predict pesticide loss for individual years, but it proved to be more reliable for the grouped years. The highest model efficiency found was 0.56. The result implies that a large part of the long-term leaching can be explained by pesticide properties.