AI-assisted metabarcoding for precise plant species identification in field surveys

Thesis Abstract

AI-assisted metabarcoding is advanced to address persistent inaccuracies in field-based plant species identification due to morphological similarities, phenotypic plasticity, and limited taxonomic expertise. The study addresses the problem of unreliable species inventories in biodiversity surveys, which impede conservation planning, invasive species management, and ecosystem monitoring. The aims are to develop a robust ICT-driven workflow that integrates metabarcoding with machine learning to yield precise species identifications from mixed-field samples, and to evaluate its performance against traditional morphological identification and standalone DNA barcoding. Specific objectives include (i) to compile and curate a reference barcode library for target taxa representative of temperate forest and grassland communities (n = 1,200 curated sequences across 300 species); (ii) to optimize an AI-assisted metabarcoding pipeline that combines high-throughput sequencing data (Illumina MiSeq, 2 × 300 bp) with convolutional neural networks and random forest classifiers to assign species with quantified confidence; (iii) to assess the pipeline's accuracy, precision, recall, and F1 scores across taxonomic ranks (species, genus) under varying sample complexity and sequencing depth; (iv) to compare performance against traditional morphological keys and a baseline DNA barcode approach using standard thresholds; and (v) to evaluate practical implications for field surveys under operational constraints. The methodology adopts a multiphase explanatory mixed-methods design. The population comprises vascular plant assemblages from three representative biomes within a temperate region mixed deciduous forests, tallgrass prairies, and riparian zones. A stratified sampling scheme yields field plots (n = 90; 30 per biome), from which bulk leaf and flower samples are collected (approximately 15–20 g per plot). Data collection involves (a) DNA extraction using a standardized plant-optimized protocol, (b) amplicon sequencing targeting a multi-locus barcode approach comprising rbcL, matK, and ITS2 regions, and (c) an auxiliary morphological inventory conducted by taxonomic specialists. The AI-assisted pipeline fuses sequence reads with metadata (habitat, phenology, and geolocation) and employs a deep learning classifier (CNN-based read-embedding) complemented by a gradient-boosted tree ensemble to predict species. Validity and reliability are addressed through technical replicates (n = 3 per plot), negative controls, and mock community standards to quantify misassignment rates. Instruments include the Illumina MiSeq platform, a validated reference database, and a custom software framework for real-time taxonomic assignment with probabilistic confidence metrics. Data analysis proceeds in three tiers. First, bioinformatic preprocessing uses established pipelines (QIIME2, DADA2) to denoise reads, generate amplicon sequence variants, and compute taxonomic abundance tables. Second, the AI-driven classification model is trained on an annotated subset (70% of reference data) and validated on the remaining 30%, with performance metrics including accuracy, precision, recall, F1, and area under the receiver operating characteristic curve (AUC). Third, comparative analyses employ McNemar tests for paired accuracy and a repeated-measures ANOVA to evaluate performance differences across biomes and sample complexity. A hierarchical Bayesian model estimates uncertainty in species assignments, integrating classifier confidence with sequencing depth. The study also examines the impact of integrating metadata on predictive performance, using ablation experiments to quantify gains from contextual attributes. Expected findings indicate that the AI-assisted metabarcoding workflow will achieve species-level accuracy above 92% for well-represented taxa and maintain robustness (80–88%) across less represented groups, outperforming conventional morphology-based identifications and baseline DNA barcoding by a margin of 12–18% in F1 scores. It is anticipated that sequencing depth saturation occurs around 25,000 reads per sample, after which marginal gains diminish. The research contributes to knowledge by demonstrating a scalable, field-applicable method that mitigates expert bottlenecks in plant identification, enhances reproducibility of biodiversity inventories, and provides a probabilistic framework for uncertainty in taxonomic calls. The findings will inform standard operating procedures for integrated DNA-based surveys and guide future expansion of reference libraries and AI models to broader biogeographic contexts. The study concludes that combining metabarcoding with AI-based classification yields reliable, rapid, and cost-effective plant identifications suitable for real-world field surveys, with recommendations for routine calibration against voucher specimens, expansion of reference datasets, and guidelines for deploying the approach in monitoring programs and conservation planning.