Forest efficiency assessment and prediction using dynamic DEA and machine learning

This paper proposes a novel Dynamic Data Envelopment Analysis (DEA) approach to assess the efficiency of forests in providing three key ecosystem services: timber production, water yield, and carbon sequestration. Carbon sequestration is modeled as a carryover (along with plot age), while timber production and water yield are considered as outputs. Given that the inputs considered (e.g. annual precipitation and average temperature, tree density, etc) are considered non-discretionary, an output orientation is used. Using a weighted additive normalized-slacks DEA model, efficiency scores are computed for each plot over the entire time horizon and for individual periods. Additionally, efficiency scores for each ecosystem service, along with corresponding slacks (e.g., carbon sequestration shortfall per hectare), are estimated. Aggregate efficiency scores for the full sample are also derived. In a second stage, regression trees (RT) and random forest (RF) models are applied to identify plot characteristics that affect ecosystem service efficiency. A case study of of 84 forest plots in Florida reveals that overall carbon sequestration efficiency exceeds timber production efficiency, with both positively correlated. Private ownership and the implementation of management practices enhance efficiency across all three ecosystem services, particularly for timber production and carbon sequestration. However, the impact of disturbances on efficiency is less clear and appears significant only within certain elevation ranges. In terms of predictive performance, RF outperforms RT in accuracy but offers lower explainability.