Reflections from the Workshop on AI-Assisted Decision-Making for Conservation
Abstract
The following is not an abstract, rather an introduction.
--------
The world has lost one-third of global forests [Williams 2008] and one million species face extinction due to human activities [Tollefson 2019]. To reverse these trends, conservation decision-makers are tasked with managing protected areas [Watson et al. 2014; Coad et al. 2015], prioritizing endangered species [Chadès et al. 2008; Schuster et al. 2019; Ball et al. 2009], and navigating opportunity cost trade-offs in conservation vs. economic development [Fox et al. 2019; Sacre et al. 2020; Dowd et al. 2022], often in the face of uncertainty, complex constraints, and long planning horizons [Visconti et al. 2010; Carvalho et al. 2017; Pressey et al. 2013; Tulloch et al. 2013]. Simultaneously, recent years have seen rapid advances in artificial intelligence (AI), including deep learning to uncover patterns in high-dimensional datasets [LeCun et al. 2015], tractable approaches to navigate exponentially large problem spaces [Browne et al. 2012], and statistical inference to learn causal effects of actions [Pearl 2019].
Much existing work at the intersection of AI and conservation focuses on extracting basic information from primary data, often in the form of computer vision for biodiversity monitoring, such as supervised learning approaches to identify animals or classify land cover [Tuia et al. 2022; Weinstein et al. 2017; Qin et al. 2022]. Gathering and interpreting this data is critical to monitor the state of the environment to inform conservation decision making. There also exist a wide range of AI methods that enable optimization, sequential planning, and inference to draw causal conclusions from observational data [CITE or forward reference sections to come].
To this end, the AI-Assisted Decision-Making for Conservation workshop convened to explore the opportunities to expand the breadth of AI methods — particularly for optimization, planning, and evaluation — that are being developed for and applied to conservation. In this report, we summarize the challenges raised during the workshop in conservation area management; methods discussed in AI, earth observation, and causality; and limitations in the application of AI for conservation.
First, we synthesize key conservation priorities. For effective conservation area management, the primary steps are to:
Understand the world. To inform conservation decision making, managers must first have a basic understanding of animal populations [CITE], species distribution [Beery et al. 2021], land cover [Qin et al. 2022], and human activity [CITE].
Act in the world. Using their best understanding of the world, managers are tasked with designating new conservationprotected areas, allocating ranger patrols, implementing human–wildlife conflict deterrence programs, intervening in the illegal wildlife trade, setting sustainable harvest targets, supporting sustainable economies, and countless other decisions — almost always in the face of limited resources and uncertainty.
Evaluate impact. Managers want to evaluate the impact of their actions, preferably to learn causal effects of these decisions, as well as better understand causal relationships between elements of the environment, such as the effect of elephant populations on carbon stocks in savannas [Davies and Asner 2019].
AI methods have the potential to aid in each of these three components of conservation management. Some approaches have already been in use for years; other approaches are not yet deployable but are growing in promise. Algorithmic approaches can help:
Understand the world. AI for earth observation can help across scales, from studying planet-level changes using remote sensing imagery [Brodick et al. 2019; Lobell et al. 2015; Hansen et al. 2013; Rolf et al. 2021; Khandelwal et al. 2022; Brown et al. 2022; Reichstein et al. 2019] to monitoring individual animal movement with photographs [Lahiri et al. 2011], acoustic monitoring, and GPS tracking.
Act in the world. AI methods can help plan across a range of contexts relevant to conservation, including sequential settings, under model uncertainty, and in the presence of multiple agents [Kochenderfer et al. 2022].
Evaluate impact. Causal inference techniques can help understand the causal effect of conservation actions, including scenarios where randomized controlled trials are too expensive or infeasible to conduct [Beyers 1998; Ferraro et al. 2018; Kimmel et al. 2021].
In this white paper, we recollect and synthesize key points made during presentations and discussions from the 2022 AI-Assisted Decision Making for Conservation workshop . In particular, we identify key open research questions in resource allocation, planning, and interventions for biodiversity conservation. WIn particular, we pinpoint conservation challenges that not only require AI solutions, but require novel methodological advances. We hope that this document serves as a call-to-action to encourage ecologists and land-useconservation decision-makers to consider adopting algorithmic decision-making approaches, and encourage AI researchers to look to conservation challenges to help prioritize which methods to advance in their future research.
The AI-Assisted Decision-Making for Conservation workshop was hosted by the Center for Research on Computation Society (CRCS) at Harvard University in October 2022. The co-authors of this paper were the workshop participants, coming from across academia, industry, and non-profits with expertise in computation, conservation, and deploying research.
--------
The world has lost one-third of global forests [Williams 2008] and one million species face extinction due to human activities [Tollefson 2019]. To reverse these trends, conservation decision-makers are tasked with managing protected areas [Watson et al. 2014; Coad et al. 2015], prioritizing endangered species [Chadès et al. 2008; Schuster et al. 2019; Ball et al. 2009], and navigating opportunity cost trade-offs in conservation vs. economic development [Fox et al. 2019; Sacre et al. 2020; Dowd et al. 2022], often in the face of uncertainty, complex constraints, and long planning horizons [Visconti et al. 2010; Carvalho et al. 2017; Pressey et al. 2013; Tulloch et al. 2013]. Simultaneously, recent years have seen rapid advances in artificial intelligence (AI), including deep learning to uncover patterns in high-dimensional datasets [LeCun et al. 2015], tractable approaches to navigate exponentially large problem spaces [Browne et al. 2012], and statistical inference to learn causal effects of actions [Pearl 2019].
Much existing work at the intersection of AI and conservation focuses on extracting basic information from primary data, often in the form of computer vision for biodiversity monitoring, such as supervised learning approaches to identify animals or classify land cover [Tuia et al. 2022; Weinstein et al. 2017; Qin et al. 2022]. Gathering and interpreting this data is critical to monitor the state of the environment to inform conservation decision making. There also exist a wide range of AI methods that enable optimization, sequential planning, and inference to draw causal conclusions from observational data [CITE or forward reference sections to come].
To this end, the AI-Assisted Decision-Making for Conservation workshop convened to explore the opportunities to expand the breadth of AI methods — particularly for optimization, planning, and evaluation — that are being developed for and applied to conservation. In this report, we summarize the challenges raised during the workshop in conservation area management; methods discussed in AI, earth observation, and causality; and limitations in the application of AI for conservation.
First, we synthesize key conservation priorities. For effective conservation area management, the primary steps are to:
Understand the world. To inform conservation decision making, managers must first have a basic understanding of animal populations [CITE], species distribution [Beery et al. 2021], land cover [Qin et al. 2022], and human activity [CITE].
Act in the world. Using their best understanding of the world, managers are tasked with designating new conservationprotected areas, allocating ranger patrols, implementing human–wildlife conflict deterrence programs, intervening in the illegal wildlife trade, setting sustainable harvest targets, supporting sustainable economies, and countless other decisions — almost always in the face of limited resources and uncertainty.
Evaluate impact. Managers want to evaluate the impact of their actions, preferably to learn causal effects of these decisions, as well as better understand causal relationships between elements of the environment, such as the effect of elephant populations on carbon stocks in savannas [Davies and Asner 2019].
AI methods have the potential to aid in each of these three components of conservation management. Some approaches have already been in use for years; other approaches are not yet deployable but are growing in promise. Algorithmic approaches can help:
Understand the world. AI for earth observation can help across scales, from studying planet-level changes using remote sensing imagery [Brodick et al. 2019; Lobell et al. 2015; Hansen et al. 2013; Rolf et al. 2021; Khandelwal et al. 2022; Brown et al. 2022; Reichstein et al. 2019] to monitoring individual animal movement with photographs [Lahiri et al. 2011], acoustic monitoring, and GPS tracking.
Act in the world. AI methods can help plan across a range of contexts relevant to conservation, including sequential settings, under model uncertainty, and in the presence of multiple agents [Kochenderfer et al. 2022].
Evaluate impact. Causal inference techniques can help understand the causal effect of conservation actions, including scenarios where randomized controlled trials are too expensive or infeasible to conduct [Beyers 1998; Ferraro et al. 2018; Kimmel et al. 2021].
In this white paper, we recollect and synthesize key points made during presentations and discussions from the 2022 AI-Assisted Decision Making for Conservation workshop . In particular, we identify key open research questions in resource allocation, planning, and interventions for biodiversity conservation. WIn particular, we pinpoint conservation challenges that not only require AI solutions, but require novel methodological advances. We hope that this document serves as a call-to-action to encourage ecologists and land-useconservation decision-makers to consider adopting algorithmic decision-making approaches, and encourage AI researchers to look to conservation challenges to help prioritize which methods to advance in their future research.
The AI-Assisted Decision-Making for Conservation workshop was hosted by the Center for Research on Computation Society (CRCS) at Harvard University in October 2022. The co-authors of this paper were the workshop participants, coming from across academia, industry, and non-profits with expertise in computation, conservation, and deploying research.
