The Promise and Peril of On-Device AI for Conservation Work

Paper Title

The Promise and Peril of On-Device AI for Conservation Work

Publication Info

• Topic area: Application of on-device AI to conservation fieldwork.
• Keywords: On-device AI, conservation technology, large language models (LLMs), EarthRanger, field data collection, speech-to-text, low-resource settings, technology acceptance, wildlife monitoring, sustainability.

Background and Problem

• Problem / challenge: Conservation fieldwork faces challenges such as data management in resource-limited settings, staffing shortages, and the need for scalable solutions. Existing AI systems often assume robust infrastructure (e.g., internet, electricity), which is not available in remote conservation areas.
• Significance: Addressing these challenges could improve biodiversity monitoring, reduce data entry burdens, and enhance the efficiency of conservation efforts in critical ecosystems.
• Motivation and related work: Advances in conservation technology have improved data collection but often exacerbate issues like high costs, reliance on external technical support, and data privacy concerns. AI assistants have shown promise in other fields (e.g., healthcare) but are underexplored in conservation, particularly for on-device deployment.

Solution

• Proposed approach: Development of an on-device transcription-language model pipeline integrated with EarthRanger, enabling field staff to dictate observations that are automatically formatted into structured data.
• Novelty:
  1. Exploration of on-device AI in low-resource conservation settings.
  2. Integration of speech-to-text and LLMs for real-time, offline data entry.
  3. Analysis of user acceptance and infrastructural challenges specific to conservation.
  4. Application of speculative methods to envision future impacts of AI deployment.
• Procedure and key techniques:
  • Conducted participant observation, surveys, and interviews with conservation organizations (Panthera and GCF).
  • Developed a prototype system combining Whisper for speech-to-text and Llama 3.2-1B for language modeling.
  • Fine-tuned the LLM using synthetic datasets derived from EarthRanger Event forms.
  • Evaluated system performance on accuracy, battery consumption, and temperature across various mobile devices.

Results

• Concrete findings:
  • Whisper transcription accuracy varied by accent, with Word Error Rates (WER) of 10.5% (Namibian-German English), 24.1% (Namibian English), and 19.6% (American English).
  • Fine-tuned LLM achieved a 96.4% field completion rate on Panthera forms, with most errors being missing fields rather than incorrect values.
  • On-device AI increased battery consumption and device temperature compared to traditional typing methods.
• Advantage over baselines:
  • Enabled offline data entry, addressing the lack of internet in remote areas.
  • Reduced manual data entry time, particularly for complex forms.
  • Improved privacy and compliance with data governance regulations by avoiding cloud-based processing.
• Experiments / evaluation:
  • Tested prototype on three mobile devices (OnePlus 7 Pro, Pixel 8, Blackview BL7000).
  • Conducted fieldwork with Panthera (cougar tracking) and GCF (giraffe surveys) to observe real-world usage.
  • Surveyed seven EarthRanger users across multiple continents on their data entry practices and attitudes toward audio-based systems.
• Limitations and future work:
  • High computational demands of on-device AI led to increased energy consumption and overheating.
  • Fine-tuning LLMs requires significant technical expertise and GPU resources, which may not be feasible for conservation organizations.
  • Future work includes deploying the system with other organizations, exploring community-based data collection, and improving model efficiency.

Summary

This paper explores the potential of on-device AI to enhance data collection in conservation fieldwork. By integrating speech-to-text and LLMs into the EarthRanger platform, the authors demonstrated that such systems could reduce data entry burdens while operating offline, a critical feature for remote settings. However, challenges such as high energy consumption, device overheating, and the need for fine-tuning infrastructure highlight the trade-offs of deploying on-device AI. The study emphasizes the importance of user-centered design and infrastructural support to ensure the sustainability and usability of AI in conservation. Future research will focus on expanding deployments and addressing technical limitations.