Irrigation ponds, or pi-tang in Chinese, are defined as an artificial construction made to impound water by constructing a dam or an embankment, or by excavating a pit or dugout. Some ponds at both microhabitat and the landscape scales may be a relevant influence for explaining bird communities due to a habitat effect or more-moderate and complex effects (Froneman et al., 2001). These ponds, regarding as wintering waterbird refuges, represent some of the multi-functional dimensions in the restoration results of agro-ecosystems. Previous studies detected that causes of species diversity are affected by habitat heterogeneity (Forman and Godron, 1986; Forman, 1995; Begon et al., 1996; Francl & Schnell, 2002; Fang et al., 2009). According to habitat selection as bio-choices, irrigation pond patterns associated with various microhabitats provide environmental clues that are used by birds to select stopover sites, such that ponds within the range of avian communities may potentially remain unoccupied or under-occupied if they lack those clues. Therefore, the appropriate microhabitats for a particular species in a guild might not be spatially constant if the habitat status changes the distance to the edge between pond cores to peripheral habitats, ie, by water-table drawdown, farmland consolidation, or other anthropogenic influences. Pond-species relationships, thus, are connected like a neural network with a non-parametric nature, as clues suggest.

In fact, estimating the avian community is a difficult task as various species may inhabit same patch in a heterogeneous landscape, so taxonomic analysis of avian guilds would be advantageously coupling them here with the development of forecasting techniques based on habitat characteristics. Surprisingly, attempts to estimate entire avian guilds with scientific rigor on such grounds are scarce in the literature, except with a few taxonomic studies (McArthur et al., 1967). Conversely, a wealth of work deals with linear predictions on a regional scale (McArthur et al., 1967; Froneman et al. 2001). In this respect, they proposed theoretical linear-relationship models using a wide range of multivariate techniques, including several methods of multivariate linear discriminant analyses, canonical analyses, and logistic regressions. Many critical reviews have indicated that these conventional models, usually based on multiple regressions, assume simple linear relationships between variables (Palmer, 1990; Reby et al., 1997). Some authors argued that regression model did not fit non-linear