A Novel Light-Informed Computational Framework for Structure-Radiation Optimization and Allocable Space Mapping in Agroforestry

Current agroforestry management systems still lack a computable and reusable framework for explicit structure-radiation evaluation, rapid optimization, and spatially explicit analysis. Existing tools provide limited support for light-environment characterization, and the integration of simulation, optimization, and usable-space mapping remains insufficient. To address this, we propose a novel light-informed computational framework based on structure-radiation coupling, including four core modules: (1) Constructing a virtual stand light-field digital twin to reconstruct three-dimensional stand scenes and designing a novel GPU-accelerated, parallel, component-wise differentiable forward radiative tracing simulator (GPU-PCDFRT) to unify structure reconstruction, quantitative light-transport solution, and radiation-response evaluation; (2) Constructing a structure-radiation surrogate model for rapid representation and repeated invocation of attribute–structure–radiation relationships; (3) Designing a light-integrated stand structure optimization agent (LiSSO-Agent) based on an Elite-enhanced Soft Actor-Critic (E-SAC) to generate selective-harvesting decisions under dual constraints of structural quality and light-environment quality; (4) Identifying, delineating, and mapping potentially allocable understorey space after harvesting to support management-oriented space identification based on post-harvest light-environment reconstruction. The framework enabled detailed three-dimensional reconstruction for 50 pure Chinese fir (Cunninghamia lanceolata) plots and effectively captured radiation distribution patterns within complex stands. The generalized additive model (GAM) trained on 1868 individual-tree samples achieved an R2 of 0.840 in ten-fold cross-validation. The E-SAC-based LiSSO-Agent improved three core optimization metrics NL, LightQ, and SR by 83.45 %, 11.33 %, and 22.58 %, respectively, with optimized NL consistently surpassing the initial state and overall outperforming heuristic baselines, and showing high stability across the 50 plots (mean CV = 0.149 %). The framework also reconstructed post-harvest understorey PAR patterns and identified potentially allocable space for the spatial arrangement of light-demanding understorey crops in agroforestry systems. Overall, this study presents a light ecological process-based optimization paradigm that explicitly integrates light into management decisions, enabling coordinated stand structure-radiation regulation and precise understorey resource allocation in agroforestry systems.