Mechanisms proposed to explain the maintenance of species diversity within ecological communities of sessile organisms include niche differentiation mediated by competitive trade-offs, frequency dependence resulting from species-specific pests, recruitment limitation due to local dispersal, and a speciationextinction dynamic equilibrium mediated by stochasticity (drift). While each of these processes, and more, have been shown to act in particular communities, much remains to be learned about their relative importance in shaping community-level patterns. We used a spatially-explicit, individual-based model to assess the effects of each of these processes on species richness, relative abundance, and spatial patterns such as the species-area curve. Our model communities had an order-of-magnitude more individuals than any previous such study, and we also developed a finite-size scaling analysis to infer the large-scale properties of these systems in order to establish the generality of our conclusions across system sizes. As expected, each mechanism can promote diversity. We found some qualitative differences in community patterns across communities in which different combinations of these mechanisms operate. Species-area curves follow a power law with short-range dispersal and a logarithmic law with global dispersal. Relativeabundance distributions are more even for systems with competitive differences and trade-offs than for those in which all species are competitively equivalent, and they are most even when frequency dependence (even if weak) is present. Overall, however, communities in which different processes operated showed surprisingly similar patterns, which suggests that the form of community-level patterns cannot in general be used to distinguish among mechanisms maintaining diversity there. Nevertheless, parameterization of models such as these from field data on the * Corresponding author. Present address: Laboratoire d'Ecologie Terrestre, Centre National de la Recherche Scientifique, UMR 5552, 13 avenue du Colonel Roche, F-31029 Toulouse cedex 4, France; e-mail: chave@cict.fr. † E-mail: helene@eno.princeton.edu. ‡ E-mail: slevin@eno.princeton.edu. strengths of the different mechanisms could yield insight into their relative roles in diversity maintenance in any given community.

No abstract

We introduce an analytical model, the Wald analytical long-distance dispersal (WALD) model, for estimating dispersal kernels of wind-dispersed seeds and their escape probability from the canopy. The model is based on simplifications to well-established three-dimensional Lagrangian stochastic approaches for turbulent scalar transport resulting in a two-parameter Wald (or inverse Gaussian) distribution. Unlike commonly used phenomenological models, WALD's parameters can be estimated from the key factors affecting wind dispersal--wind statistics, seed release height, and seed terminal velocity--determined independently of dispersal data. WALD's asymptotic power-law tail has an exponent of -3/2, a limiting value verified by a meta-analysis for a wide variety of measured dispersal kernels and larger than the exponent of the bivariate Student t-test (2Dt). We tested WALD using three dispersal data sets on forest trees, heathland shrubs, and grassland forbs and compared WALD's performance with that of other analytical mechanistic models (revised versions of the tilted Gaussian Plume model and the advection-diffusion equation), revealing fairest agreement between WALD predictions and measurements. Analytical mechanistic models, such as WALD, combine the advantages of simplicity and mechanistic understanding and are valuable tools for modeling large-scale, long-term plant population dynamics.

No abstract

One significant challenge in the construction of visual detection systems is the acquisition of sufficient labeled data. This paper describes a new technique for training visual detectors which requires only a small quantity of labeled data, and then uses unlabeled data to improve performance over time. Unsupervised improvement is based on the cotraining framework of Blum and Mitchell, in which two disparate classifiers are trained simultaneously. Unlabeled examples which are confidently labeled by one classifier are added, with labels, to the training set of the other classifier. Experiments are presented on the realistic task of automobile detection in roadway surveillance video. In this application, co-training reduces the false positive rate by a factor of 2 to 11 from the classifier trained with labeled data alone.