Ecological systems are vulnerable to irreversible change when key system properties are pushed over thresholds, resulting in the loss of resilience and the precipitation of a regime shift. Perhaps the most important of such properties in human-modified landscapes is the total amount of remnant native vegetation. In a seminal study Andrén proposed the existence of a fragmentation threshold in the total amount of remnant vegetation, below which landscape-scale connectivity is eroded and local species richness and abundance become dependent on patch size. Despite the fact that species patch-area effects have been a mainstay of conservation science there has yet to be a robust empirical evaluation of this hypothesis. Here we present and test a new conceptual model describing the mechanisms and consequences of biodiversity change in fragmented landscapes, identifying the fragmentation threshold as a first step in a positive feedback mechanism that has the capacity to impair ecological resilience, and drive a regime shift in biodiversity. The model considers that local extinction risk is defined by patch size, and immigration rates by landscape vegetation cover, and that the recovery from local species losses depends upon the landscape species pool. Using a unique dataset on the distribution of non-volant small mammals across replicate landscapes in the Atlantic forest of Brazil, we found strong evidence for our model predictions - that patch-area effects are evident only at intermediate levels of total forest cover, where landscape diversity is still high and opportunities for enhancing biodiversity through local management are greatest. Furthermore, high levels of forest loss can push native biota through an extinction filter, and result in the abrupt, landscape-wide loss of forest-specialist taxa, ecological resilience and management effectiveness. The proposed model links hitherto distinct theoretical approaches within a single framework, providing a powerful tool for analysing the potential effectiveness of management interventions.

We present a comprehensive approach to detect pattern in assemblages of plant and animal species linked by interactions such as pollination, frugivory or herbivory. Simple structural models produce gradient, compartmented or nested patterns of interaction; intermediate patterns between a gradient and compartments are possible, and nesting within compartments produces a combined model. Interaction patterns can be visualized and analyzed either as matrices, as bipartite networks or as multivariate sets through correspondence analysis. We argue that differences among patterns represent outcomes of distinct evolutionary and ecological processes in these highly diversified assemblages. Instead of choosing one model a priori, assemblages should be probed for a suite of patterns. A plant Á /pollinator assemblage exemplifies a simple nested pattern, whereas a plant Á / herbivore assemblage illustrates a compound pattern with nested structures within compartments. Compartmentation should reflect coevolutionary histories and constraints, whereas differences in species abundance or dispersal may generate nestedness.

The worldwide decline in amphibians has been attributed to several causes, especially habitat loss and disease. We identified a further factor, namely "habitat split"-defined as human-induced disconnection between habitats used by different life history stages of a species-which forces forest-associated amphibians with aquatic larvae to make risky breeding migrations between suitable aquatic and terrestrial habitats. In the Brazilian Atlantic Forest, we found that habitat split negatively affects the richness of species with aquatic larvae but not the richness of species with terrestrial development (the latter can complete their life cycle inside forest remnants). This mechanism helps to explain why species with aquatic larvae have the highest incidence of population decline. These findings reinforce the need for the conservation and restoration of riparian vegetation.