Decision-Making for Land Conservation: A Derivative-Free Optimization Framework with Nonlinear Inputs

Cassidy K. Buhler, Hande Y. BensonAAAI 2024 · DOI

Abstract

Global conservation targets like the "30x30" initiative necessitate efficient land acquisition strategies that prioritize ecological quality and connectivity. Traditional Integer Programming models often struggle with the nonlinear complexities of natural systems, such as extinction risk and habitat fragmentation. This paper presents a novel "plug-and-play" optimization framework that integrates Population Viability Analysis (PVA) tools as "black-box" inputs within a derivative-free optimization scheme. Using Extended Ant Colony Optimization (ACO), the framework identifies land conservation portfolios that maintain high species survival probabilities at significantly reduced financial costs compared to baseline "protect-all" strategies.

Introduction

The global "30x30" mandate aims to protect 30% of Earth's land and sea by 2030. However, a focus on land quantity alone may fail to ensure species persistence if habitat quality and connectivity are ignored. Traditional conservation planning relies on linear Integer Programming (IP), which oversimplifies ecological realities like species competition and extinction probability. This research bridges the gap by proposing a Mixed Integer Nonlinear Programming (MINLP) framework capable of handling complex, nonlinear ecological simulations.

Methodology

The framework treats specialized ecological software, such as RAMAS for Population Viability Analysis (PVA), as a "black box." Since these simulations do not provide gradients, the authors employ Extended Ant Colony Optimization (ACO) to navigate the search space. Two primary formulations are explored: a constrained model that minimizes cost while meeting ecological thresholds, and a multi-objective model that balances cost and ecological health. Key metrics optimized include Risk of Total Extinction, Median Time to Extinction, and Expected Minimum Abundance.

Patch identification on a 10x10 grid based on Habitat Suitability Index (HSI) thresholds.

Data and Experimental Setup

The study utilizes synthetic $10 \times 10$ and $20 \times 20$ raster grids. Each cell is assigned a Habitat Suitability Index (HSI). Contiguous parcels with HSI > 0.5 are grouped into "patches" to facilitate metapopulation simulations. This setup mimics real-world species distribution models while allowing for controlled testing of the optimization framework.

Key Results

The optimized solutions were compared against a baseline ($B$) where every habitable parcel is protected. In the $10 \times 10$ landscape, the model achieved a 0% extinction risk at a cost of 340 units—a 41% reduction from the baseline cost of 577 units. Similar efficiencies were observed in the $20 \times 20$ landscape, where costs were nearly halved (from 2137 to 1134 units) without compromising extinction risk.

Comparison of baseline (B), selected parcel status (X), and the resulting optimized conservation map (Z).

Discussion

A key advantage of this "plug-and-play" approach is its metric-agnostic nature; researchers can swap PVA tools for climate or biodiversity models seamlessly. Furthermore, connectivity is handled implicitly: fragmented habitats naturally yield higher extinction risks, leading the optimizer to favor more cohesive configurations. The results underscore that strategic parcel selection can achieve maximum ecological benefit at a fraction of the cost of total preservation.

Conclusion

While the current framework faces computational bottlenecks due to the sequential nature of PVA simulations, it demonstrates the power of derivative-free optimization in conservation planning. Future work will focus on parallelizing simulations, incorporating multi-species metrics, and validating the model with real-world geographical data.

Summary updated

BibTeX
@inproceedings{buhler2024decision,
  author = {Cassidy K. Buhler and Hande Y. Benson},
  title = {{Decision-Making for Land Conservation: A Derivative-Free Optimization Framework with Nonlinear Inputs}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i20.30195},
}
Figure