Source-linked AI summary
Species Distribution Modeling for Machine Learning Practitioners: A Review
Sara Beery, Elijah Cole, Joseph Parker, Pietro Perona, Kevin Winner
arXiv:2107.10400v1cs.LGstat.AP
TL;DR
Ecological species distribution modeling has received relatively little attention from computer science despite its relevance to conservation. This review introduces SDM concepts, models, data, and challenges, concluding with a resource for computer scientists interested in ecologically useful work.
Problem
Species distribution modeling has received relatively little attention from computer science, despite its relevance to understanding ecosystems and conservation.
Source: §1 · p. 2Method
The paper introduces SDM concepts and terminology, inventories available datasets and methods, and discusses common pitfalls and challenges.
Source: §1 · p. 3, §8 · p. 19Results
The paper provides an entry-point resource on SDM for computer scientists interested in conservation and sustainability applications.
Source: §8 · p. 19Takeaways & Limitations
The review emphasizes end-user needs and ecologically meaningful progress in machine-learning work on SDM.
Source: §1 · p. 3Long-tailed species data can leave rare-species performance poor even when metrics averaged across examples appear high.
Source: §3.5.2 · p. 12
Abstract
from arXiv · showhide
Conservation science depends on an accurate understanding of what's happening in a given ecosystem. How many species live there? What is the makeup of the population? How is that changing over time? Species Distribution Modeling (SDM) seeks to predict the spatial (and sometimes temporal) patterns of species occurrence, i.e. where a species is likely to be found. The last few years have seen a surge of interest in applying powerful machine learning tools to challenging problems in ecology. Despite its considerable importance, SDM has received relatively little attention from the computer science community. Our goal in this work is to provide computer scientists with the necessary background to read the SDM literature and develop ecologically useful ML-based SDM algorithms. In particular, we introduce key SDM concepts and terminology, review standard models, discuss data availability, and highlight technical challenges and pitfalls.
1 INTRODUCTION
Ecological studies face labor-intensive data-analysis bottlenecks, while unfamiliar ecological concepts can make SDM difficult for computer scientists to enter. This review introduces SDM and its challenges to support ecologically useful machine-learning work.
1 INTRODUCTION
- Processing ecological observations can take years for a season of camera-trap data or months to years to identify collected beetles.The paper also describes round-the-clock field monitoring of sonar data to estimate salmonid escapement.Source: §1 · p. 2
- Machine learning can accelerate analysis of large ecological data repositories, but unfamiliar ecological concepts and terminology can hinder practitioners’ entry.The paper cites progress in automating species identification from community-science and camera-trap images.Source: §1 · p. 2
- SDM uses species observations and environmental data to estimate a species’ geographic range, an area of growing interest for machine learning.The paper aims to provide computer scientists an entry point to the field.Source: §1 · p. 2
- The review highlights SDM’s technical challenges and end-user needs to encourage ecologically meaningful progress.It is intended as a quick resource for computer scientists entering conservation and sustainability applications.Source: §1 · p. 3
2 REPRESENTING THE DISTRIBUTION OF SPECIES
Species distributions can be represented using observations, statistical-model predictions, or expert knowledge. The section distinguishes these sources and notes the limits of raw observations and expert range maps.
2.1 Raw species observation data.
- Raw observation data records presence or absence, but a map of sampled locations cannot predict occurrence at unsampled locations.Whether absence data is available depends on collection strategy: opportunistic sightings are presence-only, while exhaustive checklists can support presence-absence data.Source: §2.1 · p. 3
2.2 Statistical models.
- Statistical species distribution models pair observations with environmental characteristics to predict presence or absence at places and times where those characteristics are known.Source: §2.2 · p. 3
2.3 Expert range maps.
- Expert range maps, including the IUCN caracal example [22], are often trusted but can overestimate occupancy and vary with scale.Some are refined from observation data, so they may not be independent of that data; the lack of solid ground truth also makes accuracy difficult to assess.Source: §2.3 · p. 4, Figure 2 · p. 4
3 SPECIES DISTRIBUTION MODELS
SDMs formalize species-presence prediction using environmental representations of locations. The section traces the field’s modeling history and the development of data and covariates used in those models.
3 SPECIES DISTRIBUTION MODELS
- An SDM pipeline combines presence-absence observations, environmental encodings of locations, and a parameterized model that predicts presence or absence.Source: §3 · pp. 4–5
- SDMs typically output a numerical suitability score, not necessarily a presence probability; thresholds can convert continuous predictions into binary maps.This section considers presence versus absence, while abundance prediction is treated as a more challenging related problem.Source: §3 · p. 5
3.1 A brief history of species distribution modeling
- Quantitative SDMs developed from regression and discriminant analyses to GLMs, while large presence-only datasets helped spur MaxEnt.Earlier predecessors included qualitative links between species patterns and environmental or geographic factors, such as Grinnell’s 1904 study.Source: §3.1 · p. 5
- BIOCLIM, introduced in 1984, linked observations to environmental variables and mapped possible distributions.Its climate interpolation techniques underpin WorldClim and remain widely used; modern models also draw on weather monitoring, remote sensing, elevation maps, and sparse in-situ measurements.Source: §3.1 · p. 5
- Figure 3 contrasts geographic locations with environmental-covariate space, reflecting the assumption that environmental features can support distribution mapping.Source: Figure 3 · p. 6
3.2 Covariates for species distribution modeling
SDMs use environmental covariates from climate, soils, vegetation, land cover, human influence, and remote sensing to characterize locations. These covariates vary in their sources, resolution, and suitability for species prediction.
3.2 Covariates for species distribution modeling
- Covariates span established and newer categories, and some are themselves estimated by predictive models because dense sampling of Earth's surface is costly.The section distinguishes widely used covariates from more recent or speculative ones.Source: §3.2 · p. 6
3.2.1 Climatic variables.
- WorldClim provides 19 climate-related variables averaged over 1970–2000 at around 1 km² resolution.Temperature and precipitation are identified as critical ecosystem characteristics.Source: §3.2.1 · p. 6
3.2.2 Pedologic (soil) variables.
- SoilGrids250m offers global soil properties, including pH, density, and organic carbon, at 250 m² resolution.Soil characteristics are closely related to plant life and therefore influence ecosystems.Source: §3.2.2 · p. 6
3.2.3 Vegetation indices.
- Vegetation indices are remote-sensing-derived measures of plant life; NDVI uses red and near-infrared bands and is intended to indicate live green plants.These indices can be viewed as hand-designed features for remote-sensing data.Source: §3.2.3 · pp. 6–7
3.2.4 Land use / land cover.
- Land cover describes physical terrain, while land use emphasizes function; available datasets often use coarse categories not tailored to species prediction.The U.S. National Land Cover Database assigns 20 general classes to each 30 m² patch every 2–3 years.Source: §3.2.4 · p. 7
3.2.5 Measures of human influence.
- The Human Influence Index combines eight factors, including population density, roads, nighttime lights, and land cover, into a score of human reshaping.The index is intended to quantify how much an environment has been reshaped by humans.Source: §3.2.5 · p. 7
3.2.6 Remote sensing imagery.
- Overhead imagery supplies derived covariates and may expose patterns through computer vision, but raw-imagery SDM research remains early-stage.Vegetation indices, land cover, land use, and many human-influence measures derive from overhead imagery.Source: §3.2.6 · p. 7
3.3 Properties of species distribution models
SDMs differ in their observation data, number of species modeled, treatment of location, and handling of detection and uncertainty. These categories can overlap rather than forming mutually exclusive model types.
3.3 Properties of species distribution models
- The properties used to categorize SDMs are generally nested or overlapping, not mutually exclusive.Any particular model may have some or none of the properties discussed.Source: §3.3 · p. 7
3.3.1 Presence only vs. presence-absence models.
- Presence-only data are easier to collect than presence-absence data, but limit what can be estimated; most models are designed for one type, though methods using both are emerging.The distinction depends on whether observations include information about absences.Source: §3.3.1 · p. 7
3.3.2 Single vs. multi-species models.
- Single-species SDMs model one species, whereas multi-species models capture information about several species; interest in the latter has grown.Many earlier models were single-species models.Source: §3.3.2 · p. 7
3.3.3 Multi-species models: stacked vs. joint.
- Stacked models fit a separate SDM per species and combine their maps, while joint models use other species’ occurrence data to estimate each species’ distribution.Stacking is simple but cannot exploit patterns in species co-occurrence.Source: §3.3.3 · pp. 7–8
3.3.4 Spatially explicit models.
- Environmental-feature models give identical predictions to locations with the same features, while spatially explicit models incorporate geographic location directly.This distinction matters even when locations with matching environmental characteristics are far apart.Source: §3.3.4 · pp. 8–9
3.3.5 Occupancy models.
- Occupancy models account for imperfect detection by modeling the probability that a species is present but unobserved, conditional on sampling effort.An observation confirms presence, but failure to observe a species does not prove absence.Source: §3.3.5 · p. 9
3.3.6 Understanding uncertainty and error.
- SDM uncertainty includes uncertainty in observation data and uncertainty from modeling choices, covariates, and observation locations.Prior work examines choices such as covariate selection and future-climate projections, and reviews ways to minimize or incorporate uncertainty.Source: §3.3.6 · p. 9
3.4 Algorithms for species distribution modeling
SDM algorithms have different properties, assumptions, and use cases, and are generally not readily interchangeable. The section reviews presence-only methods, traditional statistical models, and machine-learning approaches.
3.4 Algorithms for species distribution modeling
- SDM methods differ in assumptions and use cases, so they are generally not readily interchangeable.The overview follows an organization used in a comparative study of SDM techniques.Source: §3.4 · p. 9
3.4.1 Presence-only methods.
- MaxEnt estimates species-observation probability from environmental covariates, matching observed data while staying close in KL divergence to the covariate marginal distribution.Another presence-only approach generates pseudonegatives and applies traditional presence-absence methods.Source: §3.4.1 · p. 9
3.4.2 Traditional statistical methods.
- Common statistical workhorses include generalized linear models, with logistic regression and generalized additive models as special cases.HMSC relates occurrences to environmental features with a generalized linear model and can incorporate traits or evolutionary history.Source: §3.4.2 · p. 9
3.4.3 Machine learning methods.
- Machine-learning methods address complex species–environment relationships that may violate traditional assumptions, including linear covariate effects or i.i.d. sampling.Examples include boosted trees, random forests, support vector machines, neural networks, and CNNs that extract features from spatial arrays.Source: §3.4.3 · p. 10
3.5 The challenge of evaluation
Evaluating SDMs is difficult because ground truth is limited and common validation choices can misrepresent ecological performance. The section surveys evaluation options and pitfalls that affect model trust.
3.5.1 Notions of Ground Truth.
- Presence-absence surveys are an ideal comparison but are costly to repeat or expand, noisy for absences, and unavailable for most species and locations.Such evaluation remains possible for selected species and places at sparse time points.Source: §3.5.1 · p. 10, §3.5 · p. 10
- Presence-only evaluation includes false-negative rate and top-k accuracy; the former ignores other locations, while top-k requires a difficult cutoff choice.A fixed k may not fit locations with different species richness.Source: §3.5.1 · p. 10
- Community-science data, expert range maps, and downstream tasks offer alternatives, but bring sampling or validation concerns and may not measure distribution fidelity.Continuous SDM predictions and categorical expert maps also require choices about thresholds or comparison methods.Source: §3.5.1 · p. 11
3.5.2 Evaluation pitfalls.
- Random test splits can overestimate spatial-prediction performance through interpolation; hold out spatial areas for novel-location use and time for future monitoring.Randomly sampled test sets do not proxy either use case well.Source: §3.5.2 · pp. 11–12
- SDM hyperparameter selection is difficult because obtaining a useful validation set can be challenging, though spatial-autocorrelation-aware techniques have been developed.Prior work also studies sensitivity to hyperparameters.Source: §3.5.2 · p. 12
- Spatial quantization choices can produce very different outcomes, and cross-validating grid size is reliable only with enough high-quality data and may be computationally expensive.The difficulty of motivating a grid-cell size is known as the modifiable areal unit problem.Source: §3.5.2 · p. 12
- Long-tailed species data can yield poor rare-class performance despite strong aggregate metrics; per-class or class-group metrics help avoid bias toward common species.The Snapshot Serengeti dataset, for example, contains fewer than 10 gorilla images among millions collected over 11 years.Source: §3.5.2 · p. 12
3.5.3 Model trust.
- Uncertainty about where, how much, and how long an SDM is accurate contributes to ecologists’ lack of trust, while ongoing validity checks remain an open question.Needed accuracy may vary by use case and species.Source: §3.5.3 · p. 12
4 OTHER TYPES OF ECOLOGICAL MODELS
The section distinguishes SDM from mechanistic and abundance models, then introduces biodiversity measures for describing ecosystem-level patterns.
4.1 Mechanistic models
- Mechanistic models encode species-specific environmental or biological assumptions, making them difficult to scale; this paper instead focuses on correlative SDMs that need no mechanistic knowledge.An example uses plant biology to predict viable temperature ranges.Source: §4.1 · pp. 12–13
4.2 Abundance modeling
- Abundance modeling estimates the absolute or relative number of individuals at a location, extending beyond presence or absence.Source: §4.2 · p. 13
4.2.1 Population estimation.
- Population estimation counts individuals within a defined area, most often with capture-recapture methods that require distinguishing individuals.Identification has used tags or collars, and some efforts use distinctive visual features.Source: §4.2.1 · p. 13
4.2.2 Density estimation.
- Density estimation models abundance per unit area to distinguish densely from sparsely populated locations.Source: §4.2.2 · p. 13
4.2.3 Data collection procedures for abundance.
- Abundance sampling includes harvest counts, quadrats for stationary species, and line transects that record specimens intersecting a marked line.Source: §4.2.3 · p. 13
- Other abundance procedures include cue counting, distance sampling, and using environmental-DNA ratios to estimate abundance; collection methods have different innate biases.Appropriately calibrated camera traps can also use distance sampling.Source: §4.2.3 · p. 13
4.3 Biodiversity measurement and prediction
- Biodiversity metrics describe ecosystem health at a higher level, but definitions can vary; metric values may also depend on how a region is partitioned.The section assumes a partition into sub-units has been provided.Source: §4.3 · p. 14
- Species richness counts unique species; the Shannon index is entropy of the dataset’s probability distribution; and the Simpson index is the probability that two individuals drawn with replacement are the same species.Alpha diversity is average species richness across sub-units. Gamma diversity reduces to species richness at q=0 and approaches exp H(p(R)) as q approaches 1.Source: §4.3 · pp. 14–15
5 COMMON CHALLENGES AND RISKS
The section surveys practical barriers to applying machine learning in ecology, from mismatched tools and data practices to model trust, deployment, and sensitive locations.
5.1 Differences in tools
- R dominates ecology and statistics while Python dominates machine learning, limiting code and algorithm sharing; computationally demanding models may also exceed some ecologists’ resources.Source: §5.1 · p. 15
5.2 Differences in ideas and terminology
- Different concepts and terminology impede practitioners from finding and reading work across machine learning and ecology, making ecologist collaboration important for ecologically meaningful progress.The paper notes growing interdisciplinary work.Source: §5.2 · p. 15
5.3 Combining data sources
- Combining observation sources is difficult because protocols and sampling biases differ, and many projects lack accessible standardized bias descriptions or quantitative analyses.Understanding a project’s protocol may require consulting its literature.Source: §5.3 · p. 15
5.4 Black boxes, uncertainty, and interpretability
- Black-box predictions can undermine ecologists’ ability to assess influential factors, overfitting, and uncertainty, especially when models generalize across locations or time.Model outputs may inform policy, making interpretability and uncertainty quantification important.Source: §5.4 · p. 16
5.5 Norms surrounding data sharing and open sourcing in ecology
- Ecological data can be costly to collect and curate, while public datasets can support reproducibility; dataset DOIs may encourage sharing, and norms are changing.Source: §5.5 · p. 16
5.6 Model handoffs, deployment, and accessibility
- For ecological impact, models and code should be accessible and documented; deployment plans should address maintenance, and users may lack ML-package familiarity or compute resources.Source: §5.6 · p. 16
5.7 Sensitive species
- GPS obfuscation protects rare or protected species, but its adequacy for hiding photo locations is unclear and its distance may affect model accuracy and reproducibility.The text raises removing sensitive-species photos or restricting access as possible alternatives.Source: §5.7 · pp. 16–17
6 WHAT DATA IS AVAILABLE AND ACCESSIBLE?
The section inventories ecological data sources, distinguishing analysis-ready datasets from sources that must be combined, and flags limits in size, availability, and data quality.
6 WHAT DATA IS AVAILABLE AND ACCESSIBLE?
- The section separates packaged observation-and-covariate datasets from sources that can be combined to build analysis-ready datasets.Source: §6 · p. 17
6.1 Traditional analysis-ready datasets for multi-species distribution modeling
- Traditional analysis-ready datasets vary in species and covariates, and some may be too small for data-hungry machine-learning methods.The comparison described in covers five presence-absence datasets; a benchmark covers 226 species across six regions.Source: §6.1 · p. 17
6.2 Large-scale analysis-ready datasets for multi-species distribution modeling
- GeoLifeCLEF pairs observations with covariates at scale, whereas the eBird Reference Dataset combines effort-qualified checklists with land-cover and climate variables but was not maintained or publicly available as of November 2020.GeoLifeCLEF 2020 has 1.9 million observations for 31,000 species and uses presence-only observations, typically evaluated with retrieval metrics such as top-k accuracy.Source: §6.2 · p. 17
6.4 Sources for covariates
- Earth-observation covariates are available from sources including NASA, USGS, ESA, and Google Earth Engine.Source: §6.4 · p. 18
6.5 Sources for training species identification models
- Species-identification training data can come from research-grade iNaturalist datasets, biology-focused LILA.science collections, and FGVC competitions.Source: §6.5 · p. 18
7 OPEN PROBLEMS
The paper identifies open SDM problems in scaling, ecological knowledge integration, heterogeneous data fusion, and evaluation.
7 OPEN PROBLEMS
- The section presents these topics as open problems that may benefit from machine-learning tools.Source: §7 · p. 18
7.1 Scaling up, geospatially and taxonomically
- Scaling SDMs across large regions, many-species communities, and large training sets remains a challenge, including whether joint species modeling outperforms independent models.Source: §7.1 · p. 18
7.2 Incorporating ecological theory and expert knowledge
- SDMs could incorporate ecological knowledge about dispersal, community composition, range constraints, interspecies processes, phylogeny, and functional traits.Source: §7.2 · p. 18
7.3 Fusing data
- A further open problem is combining presence-only, presence-absence, abundance, and individual data while modeling processes across spatial and temporal scales.Source: §7.3 · p. 19
7.4 Evaluation
- Evaluation requires fair comparisons that account for observation bias; no single metric suffices, and useful tests may examine novel scenarios, calibration, or realistic simulations.The paper connects these questions to generalization, domain adaptation, and dataset bias and imbalance.Source: §7.4 · p. 19
8 CONCLUSION
The paper introduces machine learning researchers to species distribution modeling, highlighting its terminology and challenges while surveying available datasets and common methods.
8 CONCLUSION
- The paper introduces machine learning researchers to species distribution modeling as a challenging real-world problem domain.Source: §8 · p. 19
- It highlights common terminology, pitfalls, and challenges in species distribution modeling.Source: §8 · p. 19
- It inventories available datasets and common methods to lower the initial overhead for newcomers.Source: §8 · p. 19
- The authors hope the document will be useful to computer scientists interested in applying machine learning expertise to species distribution modeling.Source: §8 · p. 19