Practical and Ethical Issues in Big Data and Machine Learning Forecasts of Zambian Community Forestry Engagement

Abstract

Approaches integrating geospatial “big data” and machine learning will likely be increasingly used to predict conservation-related human behavior, such as patterns of local engagement, in socioecological systems. Yet, few studies evaluate both the technical and ethical aspects of such applications. Here, we provide a nation-scale worked example that combines machine learning and publicly available data to predict spatial patterns of Community Forestry establishment among 539,221 settlements across Zambia. Our model accurately predicted out-of-sample spatial establishment patterns three-quarters of the time (balanced accuracy = 76.5%, sensitivity = 64.0%, specificity = 89.1%), though it had a high false positive rate (precision = 24.3%). Accurately forecasting conservation establishment patterns for effective resource allocation requires better data on local preferences and programmatic decision-making, among other factors. Furthermore, such artificial intelligence applications risk making decision-making more technocratic, top-down, and opaque; therefore, they should only inform deliberation over possible future scenarios within wider, multistakeholder governance processes.