The modeling of ecological data that include both abiotic and biotic factors is fundamental to our understanding of ecosystems. Repositories of biodiversity data, such as GBIF , iD igBio, Atlas of Living Australia, and SNIB (Mexico's National System of Biodiversity Information), contain a great deal of information that can lead to knowledge discovery about ecosystems. However, there is a lack of tools with which to efficiently extract such knowledge. In this paper, we present SPECIES , an open, web‐based platform designed to extract implicit information contained in large scale sets of ecological data. SPECIES is based on a tested methodology, wherein the correlations of variables of arbitrary type and spatial resolution, both biotic and abiotic, discrete and continuous, may be explored from both niche and network perspectives. In distinction to other modeling systems, SPECIES is a full stack exploratory tool that integrates the three basic components: data (which is incrementally growing), a statistical modeling and analysis engine, and an interactive visualization front end. Combined, these components provide a powerful tool that may guide ecologists toward new insights. SPECIES is optimized to support fast hypothesis prototyping and testing, analyzing thousands of biotic and abiotic variables, and presenting descriptive results to the user at different levels of detail. SPECIES is an open‐access platform available online ( http://species.conabio.gob.mx ), that is, powerful, flexible, and easy to use. It allows for the exploration and incorporation of ecological data and its subsequent integration into predictive models for both potential ecological niche and geographic distribution. It also provides an ecosystemic, network‐based analysis that may guide the researcher in identifying relations between different biota, such as the relation between disease vectors and potential disease hosts.

SPECIES (Stephens et al. 2019) is a tool to explore spatial correlations in biodiversity occurrence databases. The main idea behind the SPECIES project is that the geographical correlations between the distributions of taxa records have useful information. The problem, however, is that if we have thousands of species (Mexico's National System of Biodiversity Information has records of around 70,000 species) then we have millions of potential associations, and exploring them is far from easy. Our goal with SPECIES is to facilitate the discovery and application of meaningful relations hiding in our data. The main variables in SPECIES are the geographical distributions of species occurrence records. Other types of variables, like the climatic variables from WorldClim (Hijmans et al. 2005), are explanatory data that serve for modeling. The system offers two modes of analysis. In one, the user defines a target species, and a selection of species and abiotic variables; then the system computes the spatial correlations between the target species and each of the other species and abiotic variables. The request from the user can be as small as comparing one species to another, or as large as comparing one species to all the species in the database. A user may wonder, for example, which species are usual neighbors of the jaguar, this mode could help answer this question. The second mode of analysis gives a network perspective, in it, the user defines two groups of taxa (and/or environmental variables), the output in this case is a correlation network where the weight of a link between two nodes represents the spatial correlation between the variables that the nodes represent. For example, one group of taxa could be hummingbirds (Trochilidae family) and the second flowers of the Lamiaceae family. This output would help the user analyze which pairs of hummingbird and flower are highly correlated in the database. SPECIES data architecture is optimized to support fast hypotheses prototyping and testing with the analysis of thousands of biotic and abiotic variables. It has a visualization web interface that presents descriptive results to the user at different levels of detail. The methodology in SPECIES is relatively simple, it partitions the geographical space with a regular grid and treats a species occurrence distribution as a present/not present boolean variable over the cells. Given two species (or one species and one abiotic variable) it measures if the number of co-occurrences between the two is more (or less) than expected. If it is more than expected indicates a signal of a positive relation, whereas if it is less it would be evidence of disjoint distributions. SPECIES provides an open web application programming interface (API) to request the computation of correlations and statistical dependencies between variables in the database. Users can create applications that consume this 'statistical web service' or use it directly to further analyze the results in frameworks like R or Python. The project includes an interactive web application that does exactly that: requests analysis from the web service and lets the user experiment and visually explore the results. We believe this approach can be used on one side to augment the services provided from data repositories; and on the other side, facilitate the creation of specialized applications that are clients of these services. This scheme supports big-data-driven research for a wide range of backgrounds because end users do not need to have the technical know-how nor the infrastructure to handle large databases. Currently, SPECIES hosts: all records from Mexico's National Biodiversity Information System (CONABIO 2018) and a subset of Global Biodiversity Information Facility data that covers the contiguous USA (GBIF.org 2018b) and Colombia (GBIF.org 2018a). It also includes discretizations of environmental variables from WorldClim, from the Environmental Rasters for Ecological Modeling project (Title and Bemmels 2018), from CliMond (Kriticos et al. 2012), and topographic variables (USGS EROS Center 1997b, USGS EROS Center 1997a). The long term plan, however, is to incrementally include more data, specially all data from the Global Biodiversity Information Facility. The code of the project is open source, and the repositories are available online (Front-end, Web Services Application Programming Interface, Database Building scripts). This presentation is a demonstration of SPECIES' functionality and its overall design.

The work involved in checking millions of records by hand is hard and requires thousands of human hours. At the increasing rate at which we are collecting new data from different sources with a wide range of 'quality', the problem is getting worse. An institution like CONABIO (National Commission for the Knowledge and Use of Biodiversity, Mexico) dedicates a large amount of human resources to review species records to ensure that data published by the institution has high quality. At CONABIO we are designing a system to help us direct our attention to the most problematic data. Our methodology (Stephens et al. 2019) scores a species record according to the features of its location, and it labels it as suspicious if it has a low score. A low score means that the features of the location are unusual for that species. The features of locations are the set of abiotic, like climate and topographic charactersitics and occurrences of other species in the location. Although this does not mean that a record is wrong, it may be an indicator that a record needs to be assessed. The system we are designing works in two scenarios: in one, it scores new data based on parameters adjusted from validated data; in the second, the system checks for consistency in the database, that is, it flags records of a species that seem like outliers according to the predominant records distribution for that species. Our initial tests show that we could speed up the detection process for some problematic records. In one of our tests, where we used data that were previously labeled by hand, the method flagged 624 records, out of which 70 were confirmed as incorrect data. If we look only at the precision of the results it might seem like a poor performance, however if we look at the amount of work it might save us, it looks promising because to find the same number of inaccurate records without any assistance we would have had to review almost 5,000 records. This talk is a proof of concept for this system, and details on our initial results, reviewing both weaknesses and strengths.

No abstract