Improving ecological connectivity assessments with transfer learning and function approximation

Michael Catchen, Michelle Lin, Timothée Poisot, David Rolnick, Andrew Gonzalez2023 · DOI

TL;DR

This paper proposes a two-pronged machine learning approach to break the computational and data bottlenecks in ecological connectivity modeling. By using phylogenetic transfer learning to predict species movement behavior and neural networks to approximate heavy physics-based simulations, the authors aim to enable real-time landscape restoration optimization that was previously impossible.

What Problem Does This Paper Address?

As climate change shifts habitable zones, species must move across landscapes to survive. However, human activity—urban sprawl, agriculture, and deforestation—creates "fragmentation," turning continuous habitats into isolated patches. To protect biodiversity, conservationists need to identify and restore "connectivity" corridors.

The current gold-standard tool for this is Circuitscape (CS). It treats a landscape like an electrical circuit: habitat patches are nodes, and the land between them acts as a resistor. While scientifically robust, CS faces two major "roadblocks":

  1. Data Scarcity: It requires "Resistance-to-Movement" (RTM) data—essentially a score of how much a species dislikes crossing a specific terrain (e.g., a wolf might easily cross a forest but avoid a highway). This data is missing for the vast majority of species.
  2. Computational Cost: Running these simulations on high-resolution maps is incredibly slow. Evaluating a single restoration plan can take hours, making it impossible to run the thousands of iterations required by optimization algorithms to find the best restoration strategy.

Methods

The authors propose a framework to solve both issues:

1. Phylogenetic Transfer Learning (PTL) for RTM Prediction To solve the data gap, the authors suggest that evolutionary history can predict movement behavior. If we know the RTM for a well-studied species, we can use a "phylogenetic tree" (an evolutionary family tree) and shared traits (like body mass or phenology) to predict the RTM for a related, data-poor species. This "transfers" knowledge across the tree of life.

2. CS-NN: A Function Approximator for Circuitscape To solve the speed issue, they propose replacing the heavy physics simulation with a Convolutional Neural Network (CNN).

  • Input: A three-layer "tensor" consisting of a source patch mask, a target patch mask, and the resistance surface of the landscape.
  • Architecture: A series of convolutional and pooling layers compress the landscape into a latent representation, which a multi-layer perceptron (MLP) then uses to estimate the "flow" of movement between patches.
  • Goal: Once trained on "true" Circuitscape runs, this surrogate model can provide near-instant estimates, allowing researchers to test thousands of "what-if" restoration scenarios in seconds.

Datasets

The paper outlines a pilot project focused on the Montérégie region of southern Québec. The data includes:

  • Land-cover & Projections: 30-meter resolution contemporary data and land-use projections, plus 1km resolution climate projections reaching until the year 2100.
  • Species Data: Expert-based RTM estimates for five vertebrate species used as a baseline:
  • Northern short-tailed shrew (Blarina brevicauda)
    • American marten (Martes americana)
    • Red-backed salamander (Plethodon cinereus)
    • Wood frog (Rana sylvatica)
    • Black bear (Ursus americanus)

Key Insights and Ideas

  • Connectivity as a Surrogate Modeling Problem: The most significant insight is treating ecological connectivity not just as a biological problem, but as a function approximation task. By mimicking the output of a slow, deterministic model with a fast neural net, we move from "analysis" to "optimization."
  • Evolution as a Prior: Using phylogeny as a structured way to handle missing data acknowledges that ecological traits aren't random; they are conserved through evolution. This provides a principled way to scale conservation efforts to the "thousands of species" currently ignored by specific models.
  • Graph-based Future: The authors suggest that because the output of these models is essentially a spatial graph (nodes and flows), future iterations might benefit more from Graph Neural Networks (GNNs) than standard image-based CNNs.

Key Results

As this paper is a proposal for a pilot project presented at a workshop, it focuses on the framework rather than final performance metrics. However, the authors highlight the potential for a massive shift in workflow:

  • Speedup: Moving from "minutes or hours" per simulation to near-instantaneous neural network inference.
  • Scalability: Enabling the use of optimization algorithms like simulated annealing or reinforcement learning to automatically design protected area networks.
  • Policy Impact: The work is designed to integrate directly into "Un Plan Sud pour le Québec," a multi-stakeholder agenda for land-use policy, demonstrating a clear path from ML research to real-world conservation.

Limitations and Future Work

  • Conservation of RTM: A major unknown is the "phylogenetic scale" at which movement behavior is actually conserved. It is possible that closely related species move very differently due to specialized niches, which would challenge the PTL approach.
  • Validation Challenges: Collecting real-world movement data to validate these RTM predictions remains difficult and expensive.
  • Model Architecture: The current CNN approach treats the landscape as an image, but the authors acknowledge that architectures explicitly modeling the graph structure of the environment might perform better.
  • Black-Box Approximation: While a neural net is faster, it may lose some of the physical interpretability of the original circuit-theory-based model.

Summary updated

Abstract

This is a conference paper presented at the ICLR 2023 "Machine Learning for Remote Sensing" workshop.Protecting and restoring ecological connectivity is essential to climate change adaptation, and necessary if species are to shift their geographic distributions to track their suitable climatic conditions over the coming century. Despite the increasing availability of near real-time and high resolution data for landcover change, current connectivity planning projects are hindered by the computational time required to run connectivity analyses at realistic geographic scales with realistic models of movement. This bottleneck precludes application of optimization algorithms to prioritize ecological restoration to maintain and improve connectivity. Here we propose we can make progress toward overcoming these challenges using machine-learning methods. Our proposed methods will enable rapid optimization of connectivity prioritization and extend its application to many more species than is currently possible. We conclude by illustrating how this project will contribute to efforts to apply connectivity conservation using an example of ongoing restoration in southern Québec.
BibTeX
@misc{catchen2023improving,
  author = {Michael Catchen and Michelle Lin and Timothée Poisot and David Rolnick and Andrew Gonzalez},
  title = {{Improving ecological connectivity assessments with transfer learning and function approximation}},
  year = {2023},
  doi = {10.32942/x2259c},
}
Figure