TDWG2026
Home / Talks / SAL A
talk · Tuesday 22 September · 09:09–⁠09:21 · SAL A

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

Giulio Martellucci · AI for Biodiversity Data

Recording time 28:33–41:47Open on Vimeo ↗

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

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

Overview

What this was about

Giulio Martellucci presented work from a Biodiversa+ project in which vehicle-mounted cameras captured high-resolution roadside imagery across several countries (about 2,200 km of roads, 60 target alien plant species). Applying the Pl@ntNet classifier with multi-scale tiling works but takes around 10 seconds per image, so the team used knowledge distillation from the tiling 'teacher' to a global multi-label model: first the DINOv2 vision transformer, then a ConvNeXt (DINOv3-pretrained) model that scales linearly with resolution. The ConvNeXt model was both far cheaper (about 12 hours and 2.6 kWh versus 18 days and 138 kWh for the monitored roads) and more accurate on an expert-annotated test set.

Why it matters. Transport corridors spread invasive plants; efficient models that could run in real time and be released open source make continental-scale early detection feasible.

Key ideas

In the room

  • Invasive plants are among the top five drivers of biodiversity loss; roads and railways accelerate seed dispersal.
  • Multi-scale tiling with the Pl@ntNet API (~70,000 species) gives presence/absence but is slow and costly.
  • Knowledge distillation trains a student to mimic the tiling teacher's predictions on whole high-resolution images.
  • Vision transformers scale quadratically with resolution; ConvNeXt scales linearly.
  • ConvNeXt processes about 14 images per second and uses dramatically fewer GFLOPs than tiling.
  • ConvNeXt achieved the best mAP and ROC metrics on a test set of 36 monitored species annotated by experts; tiling ranks true occurrences but produces too many false alarms.
  • Grad-CAM and dense-feature analyses show the model localises species even with occlusion and overlap.
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 · 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

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

Yuchen Zhao
talk · 22 Sept · 09:45–⁠09:57 · SAL A

Cross-Modal Coordination for Biodiversity Monitoring at Scale: Lessons from SmartWilds

Coordinating multiple sensor modalities with AI at the far edge, informed by synchronised multimodal datasets like SmartWilds, can scale and adapt biodiversity monitoring.

Jenna Kline
talk · 21 Sept · 15:45–⁠16:00 · Aula

Mainstreaming high-throughput biodiversity monitoring using AI

AI identification platforms like ARISE can take sensor data automatically to GBIF-published, standardised data for protected-area managers, but training costs are high and human verification still matters.

Niels Raes
Programme abstract

Scalable Edge AI for Real-Time Biodiversity Monitoring: A Case Study on Tracking Invasive Species in Roadside Imagery

Giulio Martellucci

Automated biodiversity monitoring is essential for tracking changes in species distributions and abundance, yet deploying cutting-edge AI on high-resolution imagery to detect diverse taxa remains challenging. While classical object detection models (e.g., YOLO) require exhaustive localized annotations that are infeasible to scale across massive taxonomies encompassing hundreds of thousands of plant species, global multi-label models offer a powerful alternative for complex species identification in a single forward pass.

In this work, we present a case study on continuous roadside monitoring to track invasive plants in near real-time. To capture localized invasions, vehicle-mounted cameras operate at high frequencies to guarantee dense spatial sampling. To develop an edge-deployable pipeline, we initially employed a multi-scale tiling framework with a Pl@ntNet model trained on global flora to automatically annotate complex roadside scenes. Through a process called knowledge distillation, these annotations are used to transfer the recognition capabilities of the heavy tiling framework (acting as a 'teacher') into a series of compact, fast models (the 'students'). These edge-deployed student models achieve performance comparable to the computationally demanding baseline across several datasets from different European countries.

We evaluate these student models on both species identification metrics and computational costs, highlighting critical operational trade-offs between image resolution, model parameters, and real-time latency. Ultimately, this framework paves the way for scalable, cost-effective ecological monitoring adaptable to diverse high-throughput data sources, including UAV surveys, plot imagery, and other extensive field campaigns.

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