Exploratory data analysis to identify factors influencing spatial distributions of weed seed banks

Abstract

Comparing distributions among fields, species, and management practices will help us understand the spatial dynamics of weed seed banks, but analyzing observational data requires nontraditional statistical methods. We used cluster analysis and classification and regression tree analysis (CART) to investigate factors that influence spatial distributions of seed banks. CART is a method for developing predictive models, but it is also used to explain variation in a response variable from a set of possible explanatory variables. With cluster analysis, we identified patterns of variation with direction of the distance over which seed bank density was correlated (range of spatial dependence) with single-species seed banks in corn. Then we predicted patterns of the seed banks with CART using field and species characteristics and seed bank density as explanatory variables. Patterns differed by magnitude of variation in the range of spatial dependence (strength of anisotropy) and direction of the maximum range. Density and type of irrigation explained the most variation in pattern. Long ranges were associated with large seed banks and stronger anisotropy with furrow than center pivot irrigation. Pattern was also explained by seed size and longevity, characteristics for natural dispersal, species, soil texture, and whether the weed was a grass or broadleaf. Significance of these factors depended on density or type of irrigation, and some patterns were predicted for more than one combination of factors. Dispersal was identified as a primary process of spatial dynamics and pattern varied for seed spread by tillage, wind, or natural dispersal. However, demographic characteristics and density were more important in this research than in previous research. Impact of these factors may have been clearer because interactions were modeled. Lack of data will be the greatest obstacle to using comparative studies and CART to understand the spatial dynamics of weed seed banks.

Characterizing complex mixed-species bird flocks using an objective method for determining species participation

“…In both classification and regression tree models (CART), each split of the target variable (e.g., branches/nodes) uses only the single most influential predictor variable (i.e., that maximizes the homogeneity within the resulting two groups) at each node (Wiles and Brodahl 2004), and this minimizes or eliminates the impact that collinearity among predictor variables can have in global regression models (Breiman et al 1984). Moreover, CART is well-suited to ecological studies involving numerous and diverse kinds of predictor variables because it lacks assumptions about underlying distributions.…”

In-Field and Soil-Related Factors that Affect the Presence and Prediction of Glyphosate-Resistant Horseweed (Conyza canadensis) Populations Collected from Indiana Soybean Fields

“…Vayssières et al (2000) also found that CART models performed better than multiple logistic regression when modeling three tree species in California. In weed science, CART analyses have been used to identify factors influencing spatial distributions of weed seed banks by Wiles and Brodahl (2004).…”

Multivariate Analysis and Data Mining of Well-Stimulation Data by Use of Classification-and-Regression Tree With Enhanced Interpretation and Prediction Capabilities

“…The CART method, however, is not a panacea for every data-mining challenge. Both strong and weak points of referred method are summarized in Table 9 (Yarus et al 2006).…”

“…CART has been applied in several areas, such as the financial industry (Cashin and Dattagupta 2008), manufacturing and marketing (Chen and Su 2008), and medical industries (Snousy et al 2011), and even in weed science (Wiles and Brodahl 2004). Different versions of decision trees have also been applied in the petroleum industry to estimate production profiles along with uncertainty assessments in long-term production forecasts (Jensen 1998); for data classification and partitioning to predict permeability from well logs (Perez et al 2005); for case-based reasoning and planning of the execution of a fracturing job (Popa and Wood 2011); to predict average production of a well from several variables, such as producer, acid volume, and strength (Yarus et al 2006); and to predict the oil production from five significant parameters (permeability, porosity, first shut-in pressure, residual oil, and water saturation) by use of a neural-based decision-tree model (Lee and Yen 2002). Recently, the boosted regression tree was applied to the data from more than 15,000 producing wells in the Barnett shale play to predict maximum gas rates and find the relative importance of the different inputs used in the treatment (Lafollette et al 2012).…”