TDWG2026
Home / Talks / SAL A
talk · Tuesday 22 September · 11:03–⁠11:15 · SAL A

Are plant image datasets ready for AI-enabled community assessment? A systematic review of bias, representativeness and annotation quality

Yuchen Zhao · AI for Biodiversity Data

Recording time 2:33:20–2:45:14Open on Vimeo ↗

Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.

The short versionCurrent plant image datasets are not yet ready for AI-enabled ecological monitoring, especially for spatially explicit tasks and annotation quality.

Overview

What this was about

Yuchen Zhao (Newcastle University) argued that ecological AI is constrained less by model architecture than by the quantity and quality of training data. She presented a four-dimension framework, inspired by clinical reporting frameworks and FAIR, for assessing plant image dataset readiness: training-relevant characteristics, image acquisition characteristics, annotation protocol, and accessibility. A systematic review of in-situ, non-crop plant object detection and segmentation datasets screened 360 unique studies and found only 40 eligible datasets, with strong geographic bias toward Europe and Asia, a dominance of agricultural applications (19 of 40 focused on wheat), neglect of background plants, and generally unsatisfactory annotation protocols.

Why it matters. Spatially explicit tasks (locating plants and estimating cover) are needed for ecological inference, and the review offers a framework for building and reporting datasets fit for them.

Key ideas

In the room

  • Imaging tasks: classification (present?), object detection (where?), segmentation (where exactly, and what cover?).
  • Framework dimensions: training-relevant characteristics; acquisition (growth stage, season, angle, location, protocol, weather); annotation protocol and QC; accessibility.
  • 508 records, 360 unique studies screened, 40 eligible datasets.
  • Object detection datasets are increasing; segmentation less so; ground-level and UAV imagery dominate with growing use of consumer devices.
  • Datasets are uneven in size and concentrated in Europe and Asia, neglecting biodiversity hotspots.
  • Wheat (19/40) and tree crowns (8) dominate; only 3 on wildflowers.
Jump in

Notable moments

In their words

Transcript

Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.

Read the transcript ↓
Loading transcript…
Keep exploring

Related talks.

See the map ↗
talk · 22 Sept · 09:09–⁠09:21 · SAL A

Scalable Edge AI for Real-Time Biodiversity Monitoring: tracking invasive plant species in roadside imagery

Distilling a tiling-based plant identifier into a single ConvNeXt model makes high-resolution roadside invasive-species monitoring fast, cheap and more accurate.

Giulio Martellucci
talk · 24 Sept · 11:03–⁠11:15 · SAL B

Toward Quantitative AI-Readiness Metrics for Biodiversity Data Infrastructures

AI readiness should be a task-specific, measurable profile rather than a single score, and current biodiversity standards lack most of the terms needed to describe it.

Yasin Bakış
talk · 24 Sept · 11:50–⁠12:05 · SAL C

From Sensor to Species: AI-Assisted Segmentation, Deep Learning Workflows, and Persistent Biodiversity Records from Earth Observation Systems

Persistent organism identifiers for remotely sensed individual trees could turn repeat Earth-observation imagery into longitudinal biodiversity records, but only if uncertainty and model provenance travel with every record.

Kit Lewers
Programme abstract

Assessing the Readiness of Plant Image Datasets for AI-enabled Spatially Explicit Ecological Monitoring: A Systematic Review of Bias, Representativeness, and Annotation Quality

Yuchen Zhao

Despite rapid advances in AI models, the applicability of ecological monitoring remains constrained by limitations in dataset quantity and quality. While most plant image datasets focus on classification for species identification, object detection and segmentation datasets remain underdeveloped. These tasks are critical for spatially explicit ecological inference, including individual localisation, abundance estimation, and vegetation structure analysis.

This review systematically identified and analysed plant image datasets developed for ecological monitoring. An assessment framework inspired by clinical image datasets evaluation was adapted to ecological monitoring requirements and applied to 42 datasets. The framework evaluated dataset scale, ecological representativeness, annotation protocols, and data management practices.

The results revealed that most datasets lacked clearly documented annotation protocols and cross-annotator consistency controls, undermining the reliability of annotations as ground truth for AI-based ecological inference. Inconsistencies in taxonomic and geographic distribution further limit both model generalisability and the applicability of image-based monitoring at global scales. Agricultural weed detection dominated current applications, while broader biodiversity monitoring remained limited.

These findings highlight that while datasets are increasingly available, their design is often not aligned with the ecological levels required for biodiversity monitoring. Both annotation quality and ecological representativeness of available datasets fall short of the requirements for robust and generalisable image-based monitoring. Future work should prioritise ecologically representative and well-documented datasets for biodiversity monitoring.

Submitted to the conference programme before the talk, so the talk itself may differ.