Title: Decision AI for Biodiversity: From Promise to Impact
Resource URL: https://www.youtube.com/watch?v=f6YdjRXF0Po
Publication Date: 2025-06-17
Format Type: Video
Reading Time: 66 minutes
Contributors: Lily Xu;Iadine Chades;
Source: AI for Good (YouTube)
Keywords: [Biodiversity Conservation, Decision Support, Markov Decision Processes, Adaptive Management, Generative AI]
Job Profiles: Operations Research Analyst;Data Scientist;Chief Sustainability Officer (CSO);Chief Strategy Officer (CSO);Machine Learning Engineer;
Synopsis: In this video, Monash University Director of the Environmental Informatics Hub and Professor of AI Iadine Chades and Oxford Postdoctoral Fellow Lily Xu explore the development and application of AI decision-making tools for biodiversity conservation.
Takeaways: [Conservation decisions are often made under uncertainty, and AI models can help optimize actions even with limited or noisy data., The urgency of biodiversity loss means that waiting for perfect information can be more harmful than making decisions with uncertainty-aware models., Decision AI needs to be interpretable and transparent to be adopted in conservation, where trust and human understanding are critical., AI for conservation must shift from isolated applications to globally coordinated frameworks aligned with biodiversity policy targets., The value of data varies, and AI can help prioritize which data to collect in resource-constrained conservation environments.]
Summary: Iadine Chades presents how decision-oriented artificial intelligence can address the accelerating biodiversity crisis, in which 25 percent of species face extinction and global species abundance has declined by 70 percent over fifty years. She distinguishes between perception AI—used for interpreting data through image recognition or natural language processing—and control AI, which supports decision making and planning under uncertainty. The heart of her work lies in applying Markov decision processes (MDPs) and partially observable Markov decision processes (POMDPs) to conservation problems, enabling the identification of optimal, sequential actions under limited budgets, dynamic ecosystems, and imperfect data.

Through a case study on British Columbia’s northern abalone and sea otters, Chades illustrates how MDPs model species interactions, management actions (such as anti-poaching or species introduction), and stochastic dynamics (including oil spills and illegal fishing). Simulations revealed that even optimal strategies may fail to achieve dual recovery targets, prompting managers to rethink assumptions and explore novel interventions. To enhance interpretability and uptake among practitioners, her team developed K-state MDPs, which compact decision trees to a user-specified number of states, and factored MDPs, which exploit causal structure for efficient representation. Both approaches produce transparent rules that guide conservation teams without overwhelming complexity.

Chades then addresses POMDPs, which account for imperfect detection of rare species and enable managers to decide when to shift resources between monitoring and active management. She demonstrates that POMDP-based adaptive management yields optimal policies even when future dynamics are uncertain, as in climate-impacted shorebird conservation. Across all methods, co-design with local agencies ensures that models reflect ground realities and earn stakeholder trust. Chades emphasizes that minimal, interpretable models often perform nearly as well as full-scale formulations and are more readily adopted by practitioners.

Looking ahead, she advocates for a paradigm shift toward a globally coordinated AI strategy aligned with the Kunming-Montreal Global Biodiversity Framework, enabling shared priorities, data standards, and open tools. She sees emerging opportunities in generative AI—such as synthesizing gray literature, extracting causal relationships, and automatically translating stakeholder objectives into decision models—but cautions that robust safeguards and interdisciplinary collaboration are essential. Chades concludes by calling for sustained investment in co-designed, explainable decision AI to help reverse the biodiversity decline with optimal use of scarce resources.
Content: ## Introduction

The accelerating loss of biodiversity poses existential risks to economies, societies, and ecosystems worldwide. In response, the United Nations International Telecommunication Union, in partnership with over forty UN agencies and co-convened by Switzerland, launched the AI for Good platform to catalyze practical AI solutions for global challenges. This session features Iadine Chades, Director of the Environmental Informatics Hub and Professor of Artificial Intelligence at Monash University, in conversation with Lily Xu, Postdoctoral Research Fellow at Oxford University and incoming Assistant Professor at Columbia University.

## The Biodiversity Crisis

Twenty-five percent of all known species face the threat of extinction, and global species abundance has plunged by seventy percent in the last half-century. Beyond its intrinsic value, biodiversity underpins services estimated to contribute forty-four trillion US dollars—over half of world GDP—to the global economy. The collapse of ecosystems imperils food security, economic stability, health, and social equity.

## Roles of AI in Biodiversity

AI applications in biodiversity fall into two broad categories:

### Perception AI

This domain encompasses techniques for interpreting environmental data, such as computer vision, acoustic species identification, and natural language processing. These tools excel at classifying images, detecting species from audio recordings, and extracting insights from textual reports.

### Decision AI

Also known as control or reinforcement learning, decision AI models support the selection of optimal actions based on current observations and predicted future states. Unlike perception AI, which primarily focuses on understanding the environment, decision AI addresses sequential decision-making under uncertainty and resource constraints.

## Decision AI Methods for Conservation

### Markov Decision Processes

An MDP defines a set of states, actions, transition probabilities, and rewards. Chades and collaborators have applied MDPs to evaluate recovery strategies for species interactions. In one study, they simulated whether British Columbia’s northern abalone and sea otter populations could simultaneously reach predefined abundance targets under various management interventions and stochastic threats.

### Partially Observable MDPs

POMDPs extend MDPs by accounting for imperfect detection and observation noise. They guide decisions on whether to invest in monitoring or direct management, thereby optimizing resource allocation for cryptic or rare species whose presence may go undetected.

## Case Studies in Conservation

#### Sea Otter and Northern Abalone

Using an MDP framework with 819 discrete states, the research team modeled species dynamics, anti-poaching patrols, translocations, and habitat risks such as oil spills. Results indicated that under even idealized conditions, simultaneous recovery of both species was unattainable without novel management approaches.

#### Shorebird Management under Climate Change

A POMDP approach enabled adaptive strategies that learn from ongoing monitoring and adjust management actions to shifting environmental scenarios. Simulation results demonstrated improved long-term outcomes by balancing surveillance and habitat interventions.

## Enhancing Interpretability

Conservation practitioners require transparent, human-readable policies. To address this, Chades’s team introduced K-state MDPs (KMDPs), which compact large decision trees into a user-specified number of states (typically below ten). They also revisited factored MDPs, leveraging causal relationships between state variables to produce concise, interpretable strategies.

## Advancing Adaptive Management

Adaptive management recognizes that managers often lack precise knowledge of ecosystem dynamics and must learn while acting. POMDP-based frameworks formalize this learning-action trade-off, providing policies that account for future uncertainty, imperfect detection, and evolving threats such as climate change.

## Towards a Coordinated Global Approach

Despite numerous proof-of-concept tools, AI for decision support remains fragmented. Chades proposes a globally coordinated initiative aligned with the Kunming-Montreal Global Biodiversity Framework to prioritize shared data standards, open-source decision models, and a unified research agenda. She further highlights opportunities in generative AI to synthesize gray literature, extract causal insights, and translate natural-language objectives into formal decision models.

## Conclusion

Decision AI has matured beyond academic exercises into practical tools for conservation under uncertainty. By combining rigorous mathematical frameworks with co-design processes, explainable policies, and minimal data requirements, practitioners can make more informed, transparent decisions. A concerted international effort is now essential to harness AI’s full potential in reversing the biodiversity crisis and ensuring sustainable ecosystem management.