Engaging Stakeholders to Identify Barriers Preventing the Use of Machine Learning in Natural Resources Management

Innovations in sensing and computation have created new opportunities for machine learning (ML) to advance natural resources management towards data-driven prediction and decision-making. Yet, the use of ML models by natural resource managers and scientists outside of research organizations remains limited. To identify barriers preventing ML from being adopted in practice, in February 2020, we convened twenty-three applied researchers and practitioners in the natural sciences who analyze data or model environmental and agricultural dynamics. These stakeholders represented a wide range of intersecting values, knowledge of ML models, sector expertise (i.e., water management, crop production, aquaculture, animal agriculture, air quality, and forestry), and organizations (i.e., federal and local government agencies, multinational companies, engineering consultancies, academia, cooperative extension). While our stakeholder workshop was designed for preliminary information gathering, we synthesized and shared important findings from the workshop to provide guidance and recommendations on how improvements in the field of ML can accelerate adoption of ML models for natural resources management. Workshop participants expressed concerns regarding the use of ML models that broadly fell into three categories: ML model transparency, availability of educational resources, and the role of process-based understanding in ML model development. Accordingly, we recommend that researchers already working at the intersection of environmental and data sciences work towards: improving ML transparency through the use of open science frameworks, avoiding framing ML models as "black boxes", developing educational resources on the use of ML models including descriptive case studies from real-world contexts, and providing guidance on how and when process-based understanding should inform ML model architecture. Further, to maximize success in transitioning ML models from theoretical to practical applications, researchers should strive to engage stakeholders early on and at regular intervals to identify emerging challenges preventing the use of ML in natural resource management decision-making.