Definitional-Closure-Constrained Multitask Retrieval of Leaf LMA, EWT, and FMC From Hyperspectral Spectra

Leaf mass per area (LMA) and equivalent water thickness (EWT) are characteristics of leaf structure and water status. The ratio of these two traits, fuel moisture content (FMC), reflects the balance between water and structural components, involving a series of ecological processes. Simultaneously retrieving these traits from hyperspectral reflectance should facilitate disentangling water and structural components, enabling per-sample physical consistency checks via definitional closure. This enhances the credibility of cross-species and cross-environment interpretations compared to traditional single-trait or stepwise retrievals. In this study, we propose a physically consistent, multitask framework based on a temporal convolutional network (TCN) that uses a shared spectral encoder to learn synergistic information across three traits. The framework uses a soft-physics formulation of "physical baseline + residual correction" to unify predictive accuracy and physical self-consistency. Experiments on the composited EcoSIS dataset demonstrate stable retrieval performance for all three traits (test-set R² ≈ 0.89–0.93) and show that 3-task joint learning significantly outperforms stepwise-based derivation (derived FMC RMSE ≈ 47.63%). Physics regularization reduces the symmetric relative difference (SRD) — a measure of internal physical inconsistency — from ~4.71% to ~0.45%, effectively mitigating the "high-accuracy illusion" of unconstrained models. The model maintains stable accuracy and consistency across spectral resolutions ranging from 1 to 50 nm. These results suggest that the proposed framework could serve as a general retrieval paradigm for composite traits with explicit analytics and provide a more interpretable and transferable solution for generalized leaf-trait retrieval across sensors and ecological domains.