Effective monitoring of forest health can be enhanced by recording the presence of a diversity of indicator species. Birds form a well-known example of such species as they are often easily identified and frequently respond to habitat change.
However, long-term monitoring of bird species usually involves expert ornithologists conducting regular surveys at remote locations, both listening to and sighting birds to confirm their presence. This is extremely expensive and time-consuming, and so there is considerable effort world-wide being put into the creation of methods to accurately identify bird species from audio recordings. We have applied modern machine-learning methods to create a classifier that analyses environmental “soundscape” recordings and predicts which birds were vocalising on each. The classifier was trained on a new gold-standard database of recordings taken at the same time as an expert survey, and have subsequently been precisely tagged by the same expert. Our classifier shows promising accuracy and we hope it will enable conservation and biodiversity monitoring efforts to be significantly improved.
We will present our classifier and share some of our results.

Organisations worldwide face the challenge of protecting threatened species with limited resources, so it is crucial that investment decisions are optimal to prevent species extinctions. Saving our Species has developed an ambitious cost-effective framework for prioritising threatened species work across over 1,000 threatened entities. Through this framework we have partnered to develop decision tools that use world-leading artificial intelligence science to maximise conservation outcomes. We will discuss how our tools address practical solutions to intractable conservation problems faced by organisations worldwide, such as where to focus limited resources across the landscape to benefit the most threatened species, or when to act and when to invest in research to fill gaps in knowledge. We will highlight the ‘stop, manage, survey’ tool that solves the problem of deciding when to invest in conserving threatened species that are hard to detect. In the absence of sightings, we could be managing locally extinct species or giving up too soon. This tool simulates a management scenario for a given species and recommends the number of years to manage for between sightings, when to survey, and when to stop investing resources where a species is unlikely to persist. The stakes are high to make our conservation dollars go further. Yet managers rarely have innovative tools to support their allocation of resources to optimise conservation outcomes. Our approaches can support organisations to make informed decisions to save species from extinction.

Spatial simulation of species populations for ecological applications is an area of research and management with broad interest, yet currently-available tools are limited by license fees, a lack of transparency and inflexibility. Presently, no single software offering allows users to develop models that incorporate common spatial and temporal factors of interest, such as climate and weather, fire, landscape dynamics, species’ physiological requirements, and evolution, in a open-source, customisable and multi-platform environment.

We believe that steps addresses this need, and demonstrate its utility using a motivating case study of greater glider populations in south-eastern Australia under realistic scenarios of habitat loss and the impacts of fires. This is but one example of how it can be used; because of its simple interface, accessible code, and flexibility to accommodate multiple data inputs and custom functions, our software actually enables researchers or environmental managers to simulate population changes for an unlimited variety of scenarios.

Modelling species distributions involves relating species occurrences to relevant environmental variables. An important step in this process is assessing how well the model predicts the distribution of a target species. We generally do this by evaluating the predictions made for a set of locations that are not included in the model fitting process -i.e. independent data.

Since fully independent data are rarely available, a common approach involves sub-sampling the data available for modelling. In ecology, this usually involves splitting data into a training set (for model fitting) and a testing set (for model validation), and this can be repeated (e.g. for cross-validation).

Cross-validation is typically done randomly. So, testing points are sometimes located close to training points. As ecological data are often autocorrelated i.e. observations close to each other (in space or time) are more similar than distant ones, random splits can lead to an over-estimation of the model’s predictive power. Spatially-separated training and testing datasets can help determine whether the model performs as well in nearby locations as it does in more distant places.

Here we introduce a recently developed tool, written in the R programming language that generates spatially or environmentally separated cross-validation folds and includes functionalities that helps modellers make informed decisions about the choice of block cross-validation techniques. The package blockCV enables ecologists to more easily implement a range of evaluation approaches that will contribute to robust estimation of the predictive performance of species distribution models.

Joint species distribution models (JSDMs) extend the standard single species distribution model (SDM) approach by allowing multiple species to be modelled simultaneously. This approach accounts for species correlations not explained by available environmental predictors. Despite increasing adoption of JSDMs in the literature, it remains unclear how JSDM predictions differ from those of standard SDMs. By stacking multiple SDMs together we can predict community assemblage or species richness, but this does not account for species correlations. JSDMs, however, allow us to perform predictions in a variety of different ways: environment-only predictions akin to the stacked SDM approach, community assemblage predictions accounting for species correlations, and species or community level predictions conditional on known occurrence states of one or more species in the community. Predictions need to be evaluated and there is a wide array of potential metrics for JSDMs and their different prediction types. These include common SDM metrics evaluated at the species level, like AUC, and metrics that operate at the community level like community dissimilarity metrics or species richness. For a case study of frog species in Melbourne, Australia, preliminary results suggest that despite likelihood-based metrics indicating JSDMs are better fit to the data than stacked SDMs, JSDM predictions accounting for species correlations are prone to overprediction. JSDMs overpredicted species richness by ~10% and had higher rates of both true and false positive predictions than stacked SDMs. The community dissimilarity metrics returned mixed results where the JSDMs simultaneously performed better and worse community assemblage predictions depending on the metric.

Comparative ecological studies often investigate the climatic, functional or morphological overlap between entities such as species or populations. We discuss the weaknesses of using existing methods to detect such overlap and present a novel application of support vector machines to the investigation of overlap in ecological space. We propose a ‘hyperoverlap’ framework to detect overlap (or non-overlap) between point data sets and present the 'Hyperoverlap' R package which implements this framework, including functions for visualisation. Hyperoverlap uses support vector machines (SVMs) to train a classifier based on ecological point data (such as climate data extracted from geographic point records) for two entities. This classifier finds the optimal boundary between the two sets of data and compares the predictions to the original labels. Misclassification is evidence of ecological overlap between the two entities. We use the global occurrence data of genera of conifers to demonstrate the theoretical and practical advantages of this method compared to existing approaches. The method proposed here is a valuable tool for studying ecological overlap in a hypervolume framework. We show that ecological overlap can be reliably detected using Hyperoverlap. We also show that our method is more stable and less likely to be affected by sampling biases than Hypervolume suite of algorithms, the approach to detecting overlap in n-dimensional space which is most comparable to our method.