A PROSPECT-guided multi-task learning hybrid framework for simultaneous retrieval of leaf chlorophyll and carotenoids

Leaf chlorophyll (Chl) and carotenoids (Cxc) jointly govern light harvesting and photoprotection. However, concurrent retrieval from leaf reflectance remains challenging because Cxc absorption is weaker and strongly overlaps with Chl in the visible range. This spectral challenge, together with the physiological coordination between Chl and Cxc, motivates a multi-task learning (MTL) formulation for joint retrieval. Here, we propose a PROSPECT-guided MTL framework for the joint retrieval of leaf Chl and Cxc from 400–911 nm reflectance. The model is trained on PROSPECT-5 simulations and adopts a shared backbone with two pigment-specific branches. It is evaluated on five in-situ datasets (ANGERS, NX, BM, XS, and JTL; n = 680), benchmarked against separate single-task networks (STL), partial least squares regression (PLSR) and PROSPECT-based lookup-table inversion (PHY). Across datasets, the proposed MTL model consistently outperforms STL, PLSR and PHY for Chl and improves Cxc retrieval in most cases. In pooled evaluation, the RMSE for Chl is reduced by 32%, 13% and 28% compared with STL, PLSR and PHY, respectively. For Cxc, the corresponding reductions are 4%, 16% and 14%. On the public ANGERS benchmark, the model achieves an RMSE of 5.56 μg/cm 2 for Chl (R 2 = 0.93) and 2.39 μg/cm 2 for Cxc (R 2 = 0.78), yielding performance competitive with prior reflectance-based reports. PROSPECT-referenced occlusion analysis further suggests that, relative to STL, MTL induces more complementary wavelength reliance between the two tasks within pigment-sensitive regions. Overall, the PROSPECT–MTL framework enables accurate and robust reflectance-only joint retrieval of Chl and Cxc across datasets. These results provide a leaf-level basis for future canopy-scale and sensor-scale extensions in pigment-informed vegetation monitoring.