A systematic review evaluating data gap-filling strategies has revealed a substantial methodological divide between ecology and data science, showing that ecological and evolutionary researchers rarely adopt advanced deep learning architectures when imputing missing species trait data. Publishing in Environmental Data Science, the authors employed the Protocol and Reporting result with Search, Appraisal, Synthesis, and Analysis framework alongside the Population, Intervention, Comparison, Outcome, and Context protocol to search the Web of Science and Scopus databases. The team screened thirteen keyword search combinations, capturing 2,387 initial records comprising 1,508 items from the Web of Science and 849 from Scopus. After applying six predefined inclusion criteria requiring English-language peer-reviewed primary research evaluating at least one imputation method on trait or general data science problems with accessible full text, the authors retained 337 focal articles, consisting of 177 ecological or evolutionary studies and 160 computational data science papers.
Narrowing their comparative evaluation to 279 studies that focused on assessing an individual imputation method, including 156 ecological or evolutionary articles and 123 data science publications, Fisher’s exact tests demonstrated an overwhelming divergence in methodological preferences between the two disciplines (p < 0.0005). Ecological and evolutionary studies were heavily dominated by conventional machine learning and evolutionary models, with random forest representing 23.7 percent of applied methods, k-nearest neighbors accounting for 16.7 percent, and phylogenetic imputation comprising 16.0 percent. In stark contrast, deep learning techniques were deployed only twice in the ecological literature, both via feed-forward neural networks, representing just 1.3 percent of total imputation applications. Conversely, computational data science exhibited a vastly more diversified profile: regression stood at 13.0 percent, k-nearest neighbors at 12.2 percent, machine learning as an explicit collective grouping accounted for 22.8 percent, and advanced deep learning frameworks constituted 20.3 percent of all utilized methods.
Temporal comparisons across four historical decadal brackets—1990 to 1999, 2000 to 2009, 2010 to 2019, and 2020 onwards—demonstrated that this methodological cleavage has persisted over time. Between 1990 and 2009, data science and ecological studies exhibited zero methodological overlap in preferred single-method approaches, while evaluations across the 2010 to 2019 and post-2020 periods confirmed statistically significant disparities through Fisher’s exact tests (p < 0.005). Within ecology and evolution, method selection remained stable between the earliest brackets (p > 0.05), but underwent a marked transition between the 2010 to 2019 interval and the post-2020 era (p < 0.0005), reflecting a clear migration from k-nearest neighbors toward random forest models. Within data science, early adoption of deep learning appeared in 2000 to 2009 through self-organizing maps and generative topographic mapping, followed by a moderate procedural shift into 2010 to 2019 (p = 0.081) and subsequent methodological stabilization through the post-2020 period (p = 0.388).
The systematic review underscored that species trait data—which the authors defined expansively to include morphological, demographic, physiological, and phenological characteristics alongside emergent community-level attributes such as mean canopy height—are severely compromised by non-random gaps. Missingness in empirical datasets typically arises via three distinct theoretical mechanisms: missing completely at random, missing at random, and missing not at random. In biological repositories, data gaps predominantly stem from four systematic origins of bias: taxonomic overrepresentation of charismatic species, geospatial accessibility favoring organisms with large geographical ranges, intrinsic sampling difficulties surrounding rare species, and observational hurdles associated with long-term traits like lifespan or dispersal. Because real-world trait datasets reflect complex missing at random and missing not at random patterns, traditional practices like listwise deletion discard substantial sample volume and inject severe bias into ecological estimations. Furthermore, simple heuristic fixes like mean imputation or last-observation-carried-forward generate substantial distortion, necessitating sophisticated computational imputation to resolve multi-trait dependencies.
A central conceptual divergence between the two fields lies in the ultimate objective of imputation and the handling of statistical uncertainty. While data science often seeks point prediction accuracy to produce a single completed dataset for predictive tasks, ecological and evolutionary investigations depend on downstream statistical inference to test biological hypotheses regarding species relationships. The systematic review highlighted the perils of the single imputation fallacy, where single completed datasets treat filled values as empirical facts, underestimating statistical variance and producing artificially narrow confidence intervals that lead to overconfident inferences. Citing recent benchmark evaluations, the authors demonstrated that single imputation frequently underperforms relative to complete case analysis in downstream regression models, whereas multiple imputation reliably recovers true parameters by drawing iterative values from conditional posterior predictive distributions and pooling results via Rubin’s rules. Consequently, any advanced machine learning or deep learning architecture considered for ecological deployment must be evaluated primarily on its capacity to propagate error through multiple imputation.
To evaluate whether neural architectures could satisfy ecological requirements, the authors compiled evidence maps detailing 96 applications of deep learning across 37 data science investigations spanning diverse data contexts. The mapped architectures included generative adversarial networks across 16 articles, feed-forward neural networks across nine papers, recurrent neural networks across nine papers, autoencoders across nine articles, hybrid models across seven publications, graph neural networks across four articles, and other architectures across 16 papers. In evaluating generative adversarial networks, the review analyzed 13 comparative studies and found that performance relative to conventional baselines was governed by the missingness mechanism. In eight of these studies—which included the only trials assessing missing at random and missing not at random scenarios, and the only tested application to trait data—adversarial networks were outperformed by conventional baselines. Conversely, the five studies reporting adversarial network superiority were conducted strictly under missing completely at random assumptions or left the missingness mechanism undefined.
Autoencoders demonstrated considerably greater promise for ecological adaptation, outperforming conventional baselines like k-nearest neighbors, random forest, and multiple imputation by chained equations in five of eight comparative studies across numerical, categorical, longitudinal, and mixed data types. In trials examining biased missing at random and missing not at random conditions, hybrid autoencoder-classifier architectures achieved top performance, while overcomplete denoising autoencoders performed comparably to conventional alternatives, although vanilla autoencoders struggled on mixed data at low missingness thresholds below 30 percent. Recurrent neural network architectures, evaluated across four comparative studies, demonstrated that frameworks such as Bidirectional Recurrent Imputation for Time Series and Long Short-Term Memory networks utilizing rectified linear unit activation functions could surpass k-nearest neighbors on longitudinal missing completely at random data, though random forest remained superior in certain comparisons. Feed-forward neural networks also showed strong performance, with multilayer perceptrons and deep belief networks outperforming decision trees and support vector machines on numerical data.
Transformer-based models demonstrated remarkable consistency across the analyzed literature, with none of the evaluated architectures performing worse than conventional baseline methods. In longitudinal trials under missing completely at random conditions, methods including Self-Attention-based Imputation for Time Series, Multivariate Time Series Imputation with Transformers, and standard transformer encoder models consistently outperformed k-nearest neighbors, with self-attention models achieving parity with random forest. However, the authors noted a substantial empirical void, as no reviewed studies tested transformers, recurrent neural networks, or feed-forward networks under biased missing at random or missing not at random mechanisms. Among alternative architectures, novel graph neural networks outperformed all competing methods on longitudinal data, and deep Gaussian mixture models exhibited robust stability across scenarios, whereas conditional score-based diffusion models and self-organizing maps suffered poor accuracy under challenging missingness conditions. Across 18 non-comparative deep learning papers, nine reported superiority over non-deep alternatives, including mixed-model forest mensuration comparisons where multilayer perceptrons surpassed mixed-effect models when clustering plots.
Bridging the gap between deep learning capabilities and ecological research demands modifying deterministic neural architectures into frameworks capable of rigorous multiple imputation and biological covariance modeling. Because many standard deep models produce single deterministic outputs, the authors highlighted several pathways for generating predictive posterior distributions, including Mixture Density Networks that predict probability distribution parameters, deep ensemble techniques that capture epistemic uncertainty across varying initializations and bootstrap resamples, and Bayesian neural networks that iteratively sample parameter weights from updated prior distributions. Furthermore, generative frameworks like variational autoencoders inherently support multiple imputation through latent sampling, while tabular prior-fitted networks like TabPFN natively approximate Bayesian inference to yield posterior predictive distributions. To ensure biological validity, deep learning models must also incorporate evolutionary history, which can be achieved by converting phylogenetic trees into tabular formats using phylogenetic eigenvector mapping. The authors emphasized that future benchmarking must abandon reliance on isolated accuracy metrics, prioritizing measures that assess the preservation of inter-trait correlations and phylogenetic signals.
To address empirical data scarcity and taxonomic biases that typically impede deep learning training within species trait databases, the review outlined innovative computational strategies that eliminate the need for massive empirical training corpora. Foundational models such as TabPFN have achieved state-of-the-art tabular classification and regression performance on modest sample sizes by utilizing in-context learning, which allows networks to adapt inference dynamically based on provided context examples without undergoing iterative parameter updates. Furthermore, transfer learning frameworks like large language model-enhanced knowledge augmentation demonstrate capability in harmonizing disparate domain data, while synthetic data generated via simulated virtual species under controlled eco-evolutionary models offers scalable, structurally sound training corpora. The authors concluded that realizing the transformative potential of deep learning in biodiversity research requires dedicated interdisciplinary collaborations between ecologists and data scientists, coupled with the establishment of standardized benchmark datasets comparable to ImageNet and CIFAR in computer vision, ensuring that advanced imputation techniques are systematically validated against realistic ecological biases.
Subject of Research: Data imputation methods for missing species trait data in ecology and evolution
Article Title: Methods for imputing trait and community data in ecological and evolutionary research
Article References: Turner, M. T., Leontidis, G., Burslem, D. F., Travis, J. M., & Henry, R. C. (2026). Methods for imputing trait and community data in ecological and evolutionary research. Environmental Data Science, 5, Article e41. https://doi.org/10.1017/eds.2026.10067
Image Credits: AI Generated
Keywords: data imputation, species traits, deep learning, ecological and evolutionary research, machine learning, missing data mechanisms, autoencoders, transformers, random forest, k-nearest neighbors, multiple imputation, phylogenetic imputation, systematic review, data science, missing at random, missing not at random, synthetic data
News Source: Blake Davidson. (November 22, 2026). Deep Learning Imputation Remains Underutilized in Ecological and Evolutionary Trait Research. Scienmag.



