Multi-source data fusion for estimating maize leaf area index over the whole growing season under different mulching and irrigation conditions

Leaf area index (LAI) is a vital indicator to identify the crop growth condition and to infer crop yield and water consumption. The generality of empirical models for estimating LAI based on vegetation indices (VI) is often questioned due to their reduced sensitivity in dense vegetation cover or the influence of external disturbances. We aimed to investigate the potential of multi-source data fusion for improving the LAI estimation accuracy in various crop growth stages (e.g., high soil back ground disturbance in seedling stage, and dense crop coverage in mid-stage). The field experiments of two varieties of spring maize under three drip irrigation levels and two film mulching conditions were conducted in the Shiyang River Basin of Northwest China in 2021 and 2022. We collected multispectral images, thermal infrared images, hyperspectral reflectance data, photosynthetically active radiation (PAR) and LAI for each year. After data processing, we first evaluated the performance of the six VIs, i.e., normalized difference vegetation index (NDVI), ratio vegetation index (RVI), difference vegetation index (DVI), nonlinear vegetation index (NLI), anti-atmospheric vegetation index (ARVI) and the optimized NLI (VI0), for LAI estimation. We then utilized partial least squares regression (PLSR) on the 2022 dataset. Inputs included canopy temperature (Tc), fraction of canopy PAR interception (FPAR), the best VI, and their products (VI*FPAR, VI*Tc, FPAR*Tc and VI*FPAR*Tc). Ultimately, the performance of the PLSR was evaluated using the 2021 dataset. The six VIs could provide accurate estimation of LAI when LAI< 4.0, while performed poorly when LAI> 4.0 (R2: 0–0.5, RMSE: 0.64–0.90 m2 m-2), indicating VI alone failed to accurately estimate LAI at dense crop cover stage. VARI was the best among the six VIs, being soil-noise resistant but highly sensitive to leaf properties. PLSR showed higher LAI prediction accuracy than the VI-based models. Especially when LAI> 4.0, the R2 of the PLSR increased by 0.23–0.46 and RMSE decreased by 0.08–0.29 m2 m-2 compared with the VI-based model. Besides, the fusion of multi-source data estimating method removed the soil background disturbance, showing high prediction accuracy at the seedling stage. These findings indicate that combining multi-source field data could reduce the saturation of dense vegetation cover and the soil noise, providing an efficient way to map the spatial and temporal distribution of LAI.