Table 1_AI-driven assessment of forest dynamics and ecosystem services in response to land use change: current progress and future perspectives.pdf

At the monitoring tier, deep learning architectures such as U-Net achieve F1-scores of 0.75–0.77 for forest change detection, with an advantage over conventional classifiers that is conditional rather than absolute, varying with label quality, spectral availability, and geographic transfer. At the assessment tier, AI has advanced carbon stock estimation through multi-source fusion, species distribution modeling in which convolutional neural networks substantially outperform Maximum Entropy models, an advantage that persists and even widens under stricter spatial cross-validation even as both methods' absolute accuracy declines, and hydrological service quantification coupled with process-based models.

At the prediction tier, machine learning-enhanced land use simulation has reached accuracies exceeding 0.93, while reinforcement learning shows potential to jointly optimize ecosystem service and economic objectives, though only at proof-of-concept stage. Persistent challenges include limited explainability, near-absent uncertainty quantification, weak cross-domain transferability, and largely sequential AI-process model integration, with direct consequences for policy applications: AI-derived carbon stock maps and ecosystem service assessments intended to inform emissions trading, conservation planning, or land use policy currently lack the quantified uncertainty and cross-region validation such applications require.

We propose a phased research agenda spanning near-term methodological consolidation, medium-term foundation model adaptation, and long-term integrated forest digital twins with equitable global coverage, offering a roadmap for translating AI and geospatial technologies into evidence-based forest ecosystem stewardship.</p>