應用two-step模式與M5’回歸樹建構台灣玉米花粉飄散預測模式及隔離距離的模式推估

Abstract

As improvements in biotechnology, the acreage and yield of genetically modified (GM) crops have been increased rapidly during the past decade. However, the influence of the GM crops for environment and human body are concerned by some researchers and customers. Therefore, the coexistence between GM crops and non-GM crops in the open field is important. Such that the farmers can choose one kind of crops to plant freely, including GM crops, non-GM crops, organic crops, or conventional crops. Moreover, customers can purchase freely the relational products. In our study, the purple-glutinous maize was used to simulate the GM crops. The non-GM crops were the white-glutinous maize. The field experiments were conducted at Puzih Branch Station in Puzih in 2009 and 2010 and Taiwan Agricultural Research Institute (TARI) in Wufeng in 2011. To construct the pollen-mediated gene flow (PMGF) models suited the climate and environment conditions in Taiwan (R.O.C), three non-linear models (exponential model, log/log model, log/square model), two-step model and M5’ regression tree were used to compare the fitting capability. Furthermore, based on the different compulsory labeling thresholds, the isolation distance calculated by the PMGF models were also discussed. Accordingly, in Puzih, the fitting ability of M5’ regression tree was better than the other PMGF models, followed by the two-step model and log/log model. The exponential model and log/square model had a poorer fitting performance. When the labeling threshold was set as 0.9%, the farthest isolation distance was 50.25 m calculated by the log/log model. However, in Wufeng, due to the wind direction, the cross pollination rates in the downwind direction were higher than those in the upwind direction. The M5’ regression tree, two-step model and log/log model had the similar performance. The results of the exponential model and log/square model were similar to those obtained from Puzih with the poor fitting performance. Moreover, in the downwind direction, the farthest isolation distance was 4.5 m calculated by the log/log model for sufficing the compulsory labeling threshold of 0.9%. In addition, the results in Puzih and Wufeng also indicated that the distance from the nearest edge of the donor field (minidist) was the first classificatiory variable in M5’ regression tree. It was also an important variable in general PMGF models, e.g. non-linear model, two-step model and MAPOD etc. Finally, the two-step model was better fitting than non-linear models and thence the isolation distance that the two-step model calculated was to conform to agricultural environment in Taiwan. M5’ regression tree is an empirical model, and the results indicate that the fitting capability can perform obviously well as increasing in the number of explanatory variable and sample information. The results indicate that two-step model and M5’ regression tree should be suggested for drafting the coexistence policy in Taiwan (R.O.C).