Introduction

Waterbird refuges represent one of the multifunctional perspectives in the restoration of agro-ecosystems. Previous studies inferred that the causes of species diversity are affected by habitat heterogeneity (Forman and Godron 1986; Forman 1995; Begon et al. 1996; Francl and Schnell 2002). Distributions of avian species within such land mosaics are correspondingly discontinuous, depending on the locations of preferred habitats, density-dependent processes, and the quality of individual patches. Thousands of farm ponds are unique landscape features of the Taoyuan Tableland in Taiwan. Farm-pond areas are dominated by fields separated by hedgerows and windbreaks. Beyond area, all parameters in pond configuration can be condensed to fall into the categories that address eco-physical identities, such as shape, depth, edge, clustering, and connectivity in a landscape scale (Forman 1995; Linton and Boulder 2000; Froneman et al. 2001; Leitão and Ahern 2002; Oertli et al. 2002). Many studies on pondscape have focussed on the aforementioned spatial configurations, such as biodiversity, pond size (PS), pond depth, pond shape, and pond sustainability (Boothby 1997; Weyrauch and Grubb 2004; Denoel and Lehmann 2006). Pondscape parameters including PS, mean shape index (MSI), mean pond fractal dimension (MPFD), etc. can be considered. To some extent, descriptive statistics were used for statistical data processing to combine parameters in spatial analysis (Halsey et al. 1986; Evertsz and Mandelbrot 1992; Gustafson and Parker 1992; Cheng and Agterberg 1995; Kronert et al. 2001; McGarigal et al. 2002). The study determines whether there is a relationship between the PSs and numbers of species and individuals, the richness, and diversity determined by standardized sampling units.

This study assesses a non-linear relationship using neural network models instead of linear regression. Artificial neural network (ANN), which originated about several decades ago (McCulloch and Pitts 1943), was inspired by a desire to emulate human learning. ANN is highly effective for modeling nonlinear problems. The practical implication is that an ANN can accurately predict nest occurrence and breeding success of red-winged blackbird in response to ecological applications (Ozesmi et al. 2006). They developed ANN model from data in two regions and years apart for a marsh-nesting bird, the red-winged blackbird Agelaius phoeniceus. The first model was developed to predict the occurrence of nests in two wetlands on Lake Erie, Ohio in 1995 and 1996. The second model was developed to predict breeding success in two marshes in Connecticut, USA in 1969 and 1970. Independent variables were vegetation durability, stem density, stem/nest height, distance to open water, distance to edge, and water depth. Results showed that the ANN model improved predictive performance over the generalized linear models. This study attempts to investigate bird diversity and its relationship with pond attributes using ANN approach. The results will be useful in ecological planning to solve these aforementioned problems created by anthropogenic driving forces. Hence, the study objective aims at characterizing the diversity of bird species associated with these ponds. Such characterization helps to establish decision criteria needed for designating certain ponds for habitat preservation and developing their protection strategies.

Materials and methods

Study areas

Taoyuan Tableland irrigation ponds are ecologically significant because one-fifth of all the bird species find home on these ponds (Chen 2000; Fang 2004a). This tableland, at an area of 757 km2 in size, comprises an area of 2,898 ha of irrigation ponds on the northwestern portion of Taiwan. Located approximately 30 km from the capital city of Taipei, this rural area was easily converted to urban lands due to the aggregated effects of urbanization and commercialization. Socioeconomic benefits are driving public opinion, which is urging the government to approve land-use conversion from farmlands into urban uses. The Taoyuan Tableland lies between the northern border of the Linkou Tableland (23°05′N, 121°17′E) and the southern border of the Hukou Tableland (22°55′N, 121°05′E); it borders the town of Yinge in the east (22°56′N, 121°20′E) and the Taiwan Strait in the west (22°75′N, 120°99′E) (Department of Land Administration 2002) (Fig. 1). It sits at elevations from sea level to 400 m and is composed of tableland up to 303 m and hills with sloping gradients from 303 to 400 m. It runs in a southeast-to-northwest trend, abutting mountains in the southeastern corner and the shore of the Taiwan Strait at the far end. With a high average humidity of 89%, the tableland is located in a subtropical monsoon region with humid winters and warm summers. January temperatures average 13°C, and July temperatures average 28°C. Annual average precipitation ranges from 1,500 to 2,000 mm.

Fig. 1
figure 1

Spatial distribution of irrigation ponds and land use pattern in the Taoyuan Tableland of Taiwan

Full size image

The tableland gradually rose approximately 180,000 years ago. At that time, the Tanshui River had not yet captured the flow from the ancient Shihmen Creek, which directly poured out of the northwestern coast forming alluvial fans. Eventually, foothill faults caused by earthquakes during the same era, resulted in the northern region of Taiwan abruptly dropping by 200 m, and thus, the Taipei basin was born. Since the Taipei area had subsided, the ancient Shihmen Creek, which meandered across the Taoyuan Tableland was captured by northward-flowing rivers some 30,000 years ago. The middle streams changed their courses because of the subsidence in the Taipei basin. The resulting Tahan Creek, became the upstream portion of the Tanshui River in the Taipei Basin. Due to blockage of water sources, downstream areas on the Taoyuan Tableland were deficient in water. This caused high flushing and drops in water yields. Historically, it was difficult to withdraw and supply irrigated surface water from rivers due to the tableland’s unique topography, thus, forming an obstacle for the development of agriculture (Huang 1999; Chen 2000).

This area has a population density of 2,331 persons/km2 and its population is increasing at a rate of 2,000–3,000 per month. Population pressures have contributed to reductions in historical areas of farmlands and irrigation ponds (Fang 2001). Losses of farm-pond and farmland habitats have had series effects on a range of avian communities as well as other fauna and flora (Fang and Chang 2004). On the Taoyuan Tableland, agricultural practices are intensifying, which is reducing the heterogeneity of the existing landform, and adding pollutants, also resulting from industrial practices.

Survey method

This survey was based on an intensive 4-month bird survey in which simultaneous censuses were carried out at 45 ponds four times from November 2003 to February 2004 (pond plots see Fig. 1). All surveys were conducted by 45 experienced bird observers starting at the same time before sunrise and ending at 10:00 am on the same day. Each pond was surveyed and coded for numbers of bird species and individuals within 30 min with a point-count approach (Bookhout 1996). Figure 1 also shows the land-use patterns including farm ponds, built-up areas, and roads. The surrounding area at a 564.19-m basal radius from the pond geometric center (comprising a 100-ha circle) was surveyed by line-transect methodology, as shown in Fig. 2. Since the land-use ratio of basal areas (e.g., %FARM, %BUILD) within the 100-ha circle was calculated, the area with structures was used to infer attributes unsuitable for wetland birds (i.e., waterfowl, shorebirds, birds of the water’s edge, etc.) To reduce the effects of bird-observer identified bias, three to four observers were grouped and rotated between ponds. The observers counted birds that were in any habitat. Birds belonging to the families Apodidae (swifts) and Hirundinidae (swallows) were also included from counts in flight. The Shannon–Wiener diversity index (H′) was calculated to determine bird diversity, and results are discussed in the following sections. The proportion of avian species (i) relative to the total number of species (P i ) was calculated, and then multiplied by the logarithm of this proportion (log2 P i ). The resulting product was summed across species, and multiplied by −1:

Fig. 2
figure 2

Studied circles with a 564.19-m radius to delineate an area of 100 ha

Full size image
$$ H^{\prime} = - \sum\limits_{i = 1}^{S} {P_{i} } \,\log_{2} \,P_{i} $$
(1)

where S is avian species richness and P i is the percentage of species i in the avian community.

The habitat variables such as PS, pond shape, proportion of farmland area in peripheral areas, and proportion of areas with structures in peripheral areas were identified based on field surveys and Taiwan’s Geographic Aerial Map at a 1:5,000 scale (Department of Land Administration, Ministry of the Interior 2002) and Aerial Photographs at a 1:5,000 scale of 2003 (Agricultural and Forestry Aerial Survey Institute 2003).

Modeling formation

The ANN model was applied to the avian assemblage of the Taoyuan Tableland, Taiwan. One biological variable was selected to describe its structure: Shannon–Wiener’s diversity index (H′) of the waterbirds. Four habitat variables were selected as explanatory parameters: PS, MPFD, proportion of farmland area in peripheral areas (%FARM), and proportion of areas with structures in peripheral areas (%BUILD). Table 1 summarizes the descriptive statistics of habitat variables such as mean, standard deviation (SD), median (Med), etc. The comparison reveals that PS varies sharply and MPFD varies between 1.22 and 1.51. Ponds have the high values of fractal dimension which shows that the edge is more complex. When the patch’s fractal increases, movements of biological species along the margin increases; on the contrary, when the fractal on the margin decreases, biological species might cross the margin. Moreover, the land uses in the areas are approximately 72.5 and 11.4% for farmland and built-up land. These areas have a large percentage of the farmland associated with green space that is strongly correlated with waterside bird abundance.

Table 1 Descriptive statistics for four habitat variables
Full size table

Figure 3 shows the back-propagation (BP) neural network architecture which consists of four layers of neurons connected by weightings. The ANN models are developed from 35 of the 45 sampled irrigation ponds chosen at random and were validated using the 10 remaining sampled irrigation ponds. The role of each parameter was evaluated by inputting fictitious configurations of independent parameters and by checking the response of the model. The information was captured by the network when input data passed through the hidden layer of neurons to the output layer. The weightings connecting from neuron i to neuron j were denoted as w ji . The output of each neuron was calculated based on the amount of stimulation it received from the given input vector, x i , where x i was the input of neuron i. The net input of a neuron was calculated as the weights of its inputs, and the output of the neuron was based on some sigmoid function, which indicated the magnitude of this net input. So the net output, y j , from a neuron can be indicated as

Fig. 3
figure 3

Structure of the neural network used in this study

Full size image
$$ u_{j} = \sum\limits_{i = 1}^{P} {(w_{ji} x_{i} - \theta_{j} )} $$
(2)

And

$$ y_{j} = \varphi (u_{j} ) $$
(3)

where w ji is the weighting from x i to u j , x i is the ith input i.e. the habitat variable, y j is the output of jth neuron, i.e. waterbird diversity, u j is the jth neuron from an outgoing signal to the magnitude of all observations, θ j is a threshold to be passed through by the non-linear activation function, φ(·), φ(u j ) is the activation function of u j , such as the sigmoid function, and p is the number of inputs.

For this research, we chose the continuous sigmoid as the basic function:

$$ \varphi (u_{j} ) = {\frac{1}{{1 + \exp ( - u_{j} )}}}.$$
(4)

Results and discussion

Avian grouping

According to the recordings for 94 species, 15,053 individuals were detected in 4 months, microhabitats were categorized into seven guilds: air feeders (10 species), waterfowl (9 species), shorebirds (14 species), waterside birds (22 species), woodland birds (20 species), scrubland birds (13 species), and grassland birds (6 species). The value of dissimilarities was divided in accordance with distance (marked at a distance of 0.25) into 7 guilds. If this classification was adopted, the low similarities (marked at a distance of 0.75) could be divided into four guilds: waterfowl (nine species), shorebirds (14 species), waterside birds (22 species), and land birds (49 species), respectively. The likelihood of species occurrence was surveyed and categorized into a concentric pattern, such that the gradients run from the pond’s core to its edge: (1) interior pond species comprised of waterfowl (families Anatidae and Podicipedidae) and shorebirds (families Charadriidae and Scolopacidae); (2) waterside species (families Ardeidae, etc.); and (3) external pond species of land birds (i.e., species detected in such microhabitats as grasslands, scrublands, and woodlands; of families Corvidae, Frigillidae, Laniidae, Passeridae, Pycnonodidae, Sylviidae, and Zosteropidae, etc.) which were dominant in their respective microhabitats.

We then proposed that waterfowl, shorebirds, and waterside birds (families Anatidae, Charadriidae, Scolopacidae, and Ardeidae) in microhabitats were associated with distribution patterns of “interior species–edge species”. Therefore, the pond areas mainly provide shelter and foraging for wintering migrants and residents. Microhabitat structural influences on bird distribution patterns can be classified into pond core, edge, and surrounding land uses (Fang 2004b; Fang et al. 2004). On irrigation ponds, we detected more than half of the species richness of land birds, and species of land birds did not change much among the 4 months (Table 2). To relate the richness, abundance, and diversity of guilds to habitat variables, correlation analyses were carried out. The species–pondscape relationships were investigated between ecological groups and areas surrounding the ponds. Such analyses restricted to pooled guild data to reduce confounding area effects. The parameters found to be important as local determinants of community structures are the amount of farmland and amount of urban environment. Specific land uses found to be important were low-rise residential houses and high-density apartments.

Table 2 Number individuals detected and species richness of avian guilds in irrigation ponds of Taoyuan, Taiwan
Full size table

Modeling application and discussion

The resulting habitat profiles depict the complex influence of each habitat variables on the biological variables of the assemblage, and the non-linear relationships. The four habitat variables were selected as the inputs: PS, MPFD, proportion of farmland area (%FARM), and proportion of areas with structures (%BUILD) in peripheral areas. The diversity of waterbirds (H′) was predicted throughout the exercise using the back-propagation (BP) algorithm with a three-layered neural network. The first layer, called the input layer, was composed of four cells representing each of the habitat variables. The second layer, or hidden layer, was composed of a further set of neurons whose number depends on the best-calculated unbiased results. Since the BP algorithm was trained by the least mean square method, the least mean square training could reduce the error or distance between the actual output and the desired output, by adjusting the weightings. Training cases are presented sequentially, and the weightings were adjusted. We determined the number of second-layer neurons through a series of interactions which varied from two and four to eight neurons. We calculated the correlation coefficients between true values of H′ and the predicted value of H′ from the ANN. We determined the uses of non-linear models by ANN, of the number in the hidden layer for four neurons, without choosing two or eight neurons. Correlations between observed values and values estimated by ANN models of the four dependent parameters were moderately significant. The correlation coefficient (r) for four neurons was detected to differ between the training set (r = 0.726, n = 35) and the validation set (r = 0.723, n = 10) in contrast to a lower information status from a linear regression model (r < 0.28). Results showed that the ANN model is effective for predictive performance over the linear regression.

In this study, ANN predicts the diversity in the sampling sites and then inverse distance weighting interpolations are performed based on the ANN results. Figure 4 shows the resulting map of waterbird diversity in the winter of 2003. The black represents the irrigation ponds. The ponds are important landscape features that provide habitat for both breeding and migrating birds (Paracuellos and Telleria 2004; Paracuellos 2006; Hanowski et al. 2006). These ponds may be important as local food sources for both breeding and migrating individuals due to the presence of emerging insects from the ponds. Results also show that the diversity of waterbirds in western Taoyuan is higher than that of the east part, especially in the southwest part on the map (Fig. 4). Since Eastern Taoyuan is more urbanized than the southwest area (Fig. 1), the habitat in southwest of Taoyuan is better than that of the east part. Accordingly, the species of waterbirds are few in the vicinity of the built-up land. Thus, the waterbird diversity is low in the eastern Taoyuan. The presence of adjoining natural and farmland habitats was probably the most important determinant of the avifauna in farm-pond areas. We compared the following characteristics against the corresponding ratio of constructed area values associated with pond configurations at each site (Fig. 1). Evidence for an effect of human disturbance on birds is generally assumed that the presence of human infrastructures affects the distribution negatively. Due to the transportation, and urban development in Taoyuan County, farm ponds and farmland have continued to decrease. The increase of constructed areas decreases the species of birds (Marzluff and Ewing 2001). The further allocation to the current inappropriate land uses for urban development in the surrounding farm-pond areas might pose a threat to waterside bird group (Marzluff and Ewing 2001; Osborne et al. 2001). Moreover, the pond areas help wintering bird refuges and the diversity of waterbirds correlates with the pond parameters such as the number of ponds and PSs. Previous studies have shown that the farmlands associated with green space, which might translate to greater insect abundance, were strongly correlated with waterbird abundance (Fang 2005). Results from the same study also imply that the waterbirds are affected by the pond pattern at large spatial scales. As the ponds cluster together, the likelihood of discovering new creatures increase, thereby increase its species diversity.

Fig. 4
figure 4

Waterbird diversity map in the study area

Full size image

Regarding this detailed study, there may be a number of reasons why few irrigation ponds do not become refuges for waterbird species. First, there is very little of the ornamental vegetation cover found on the surrounding areas, and it may support only a small insect population. Second, the small PS associated with a curvilinear shape is not optimal to support, preserve, and attract waterbirds and other avifauna because of lacking of the management practices to support food chains. Third, anthropogenic structures are usually made of concrete with no native trees, and this may make such areas unattractive to waterbird species that require an intact shrub layer, dead wood, or generally undisturbed microhabitats.

Conclusion

The study applied ANN to investigate the relationship between the habitat variables and the waterbird diversity. The results indicate that the predictive performance of the model is better than the output of the regression model, confirming the non-linearity of the relationship among the variables. The habitat variables such as the PS, MPFD, proportion of farmland area, and proportion of areas with build-up in peripheral areas are good predictors of waterbird diversity. This analysis can help the agencies responsible for planning pondscape management to better promote use of ecological simulation skills and, thereby, to assess bird diversity in farm-pond areas.

In addition, integrated pondscape research must be a long-term project. It is necessary to increase time frame for monitoring. For example, to undertake evaluations of pondscape change requires avian data of long-term losses or increases of richness and abundance from year to year. This consideration would helps to make the simulation and evaluation model more precise in continuing this work over the long run.