Improving biodiversity protection through artificial intelligence

Biodiversity, encompassing the variety of life on Earth, provides essential ecosystem services and supports human well-being. However, unsustainable practices are driving mass extinctions, threatening the planet's life-sustaining systems. The failure to meet the Aichi Biodiversity Targets underscores the urgent need for more effective conservation policies. Spatial conservation prioritization (SCP) has emerged as a crucial step, focusing on identifying areas for protection and restoration. Existing tools, such as Marxan, while useful, have limitations: they are often designed for one-time policies, don't directly incorporate changes over time, and assume a single initial data gathering. This study addresses these limitations by leveraging the power of artificial intelligence (AI) to optimize biodiversity protection in a dynamic world.

Since the 1960s, various theoretical and practical frameworks have underpinned biological conservation. Initially focusing on nature's intrinsic value, the field evolved to acknowledge the bidirectional links between humans and nature, including the sustainable use of species. SCP and systematic conservation planning emerged as critical components, aiming to identify priority areas for protection. Several tools and algorithms have been developed, often exploring trade-offs between variables, offering significant economic, social, and environmental gains. Marxan, a widely used method, identifies a set of protected areas that meet conservation targets at minimal cost using simulated annealing. However, Marxan and similar methods lack the capacity to incorporate temporal dynamics, changes in anthropogenic pressure, and species-specific sensitivities to these changes. AI solutions have been proposed in conservation science, but reinforcement learning (RL) has not yet been implemented in practical conservation tools.

This study developed CAPTAIN (Conservation Area Prioritization Through Artificial Intelligence), a novel tool for systematic conservation planning. CAPTAIN uses a reinforcement learning (RL) framework based on a spatially explicit simulation of biodiversity and its evolution over time. The RL algorithm balances data generation ('exploration') and action ('exploitation'), quantifying outcomes ('reward'). The platform enables the assessment of model assumptions and optimizes both static and dynamic conservation policies. Actions are decided based on the system's state through a neural network whose parameters are optimized to maximize the reward. Once trained, the model can be used with simulated or empirical data. The simulation framework models biodiversity loss, considering natural processes, anthropogenic pressures (habitat modification, selective removal of species), and climate change. Anthropogenic disturbance alters natural mortality rates and carrying capacity, while climate change affects mortality based on species-specific climatic tolerances. The model optimizes two types of actions: (1) monitoring, which provides information about the system's biodiversity state, and (2) protecting, which selects areas for protection from disturbance. Different monitoring strategies (full recurrent, citizen science recurrent, and full initial monitoring) were explored, varying in detail and temporal resolution. The protection policy, informed by monitoring, selects protected areas, considering a limited budget and the cost of protection (which can vary spatially and temporally). The RL algorithm optimizes the protection policy to minimize species loss or other defined objectives (e.g., maximizing protected area or minimizing value loss). The protection policy is implemented as a feed-forward neural network, with parameters optimized to maximize the expected reward. The algorithm uses a combination of genetic strategies and policy gradient methods, allowing for parallelization and efficient parameter updates. Comparisons were made against Marxan, a state-of-the-art conservation planning tool. Empirical validation was performed using a Madagascar biodiversity dataset of endemic trees, with cost of protection proportional to anthropogenic disturbance.

CAPTAIN consistently outperformed existing methods in simulations and empirical analyses. Full recurrent monitoring resulted in the smallest species loss (26% more species protected than a random policy). Citizen science recurrent monitoring performed similarly well (24.9% improvement). Policies minimizing species loss based on commercial value sacrificed species richness for high-value ones, resulting in only 10.9% fewer species lost compared to random baseline. Maximizing protected area led to substantial species loss (13.6% more species lost than random baseline). Species that went extinct in simulations were characterized by small initial ranges, small populations, and intermediate or low resilience to disturbance. Surviving species had either low resilience but widespread ranges, or high resilience with small ranges. CAPTAIN's optimized policies protected a diverse range of species assemblages, not just areas of highest species richness. Benchmarking against Marxan revealed that CAPTAIN consistently outperformed Marxan in simulations (64% better in one-time protection, 77.2% better in dynamic protection scenarios). In the empirical analysis of Madagascar endemic trees, CAPTAIN solutions met the target of protecting 10% of species’ potential range in 68% of replicates compared with 2% for Marxan, with a significantly higher median protected range (22% vs. 14%). CAPTAIN provided more interpretable prioritization maps at higher spatial resolution.

CAPTAIN's superior performance highlights the benefits of integrating RL into SCP. The finding that even simple presence/absence data, potentially obtainable through citizen science, can inform effective policies is significant, emphasizing the potential for cost-effective conservation strategies. The model's ability to incorporate temporal dynamics is critical because many systems change drastically over time in response to climatic and anthropogenic pressures. The trade-off analyses demonstrated the importance of prioritizing biodiversity protection, not simply maximizing protected area or economic value. Contrary to intuition, the optimal strategy was not to exclusively protect the richest sites; rather, protection should span areas with intermediate to high species richness to enhance complementarity for multiple species. The successful application of a model trained using simulated data to empirical data suggests robustness and broad applicability. Future improvements could be achieved using more sophisticated features extracted from the system, including functional diversity and more detailed economic values.

CAPTAIN offers a promising tool for informing conservation decisions, outperforming existing methods in both simulations and real-world applications. Its capacity to integrate temporal dynamics and multiple biodiversity metrics is crucial in a rapidly changing world. Future work could focus on incorporating additional variables (e.g., functional diversity, carbon sequestration), refining the model through transfer learning with empirical datasets, and exploring optimal adaptive management strategies within the reinforcement learning framework.

The accuracy of the simulations depends on the quality and completeness of the input data. The model's assumptions (e.g., species' responses to disturbance, climate change) could be refined with more detailed data. While the model considers multiple biodiversity metrics, the weighting of these metrics might need to be adjusted depending on conservation priorities. The empirical application focused on a single dataset; broader validation across different ecosystems and taxonomic groups is needed.