Enhancing Leaf Nitrogen Estimation From Satellite Multispectral Imagery via Hyperspectral Super-Resolution and Structural Correction

Leaf nitrogen content (LNC) is a key regulator of photosynthesis and crop productivity, and its timely monitoring underpins both precision nitrogen management and large-scale ecosystem assessment. Remote sensing offers non-destructive, spatially explicit monitoring of LNC, but retrieval accuracy is degraded by canopy structural effects and the limited spectral resolution of multispectral satellite sensors. Here we develop an integrated framework that combines spectral super-resolution with canopy structure correction to improve satellite-based LNC estimation. Sentinel-2 multispectral reflectance is first reconstructed into full hyperspectral spectra using a probabilistic diffusion model (DDPM). Directional area scattering factor (DASF) is then computed from the reconstructed spectra to derive the canopy scattering coefficient (CSC), which normalizes structural effects and serves as input to a random forest regression (RFR) model for LNC estimation. On simulated data, the DDPM achieves high reconstruction fidelity (PSNR = 41.49 dB, SAM = 0.9989). Using CSC derived from DDPM outputs improves LNC estimation compared with multispectral bidirectional reflectance factor (BRF) alone (R² = 0.579, RMSE = 0.431 %, RRMSE = 23.58%; RMSE reduced by 14.8%, R² increased by 0.134). Even when only RGB bands are available, the super-resolution + CSC pipeline increases LNC retrieval accuracy (R² +0.14), and the addition of an NIR band further enhances both reconstruction quality and LNC estimation. This study shows that hyperspectral super-resolution coupled with structural normalization can unlock more accurate, scalable LNC monitoring from existing satellite archives, supporting operational nitrogen management and ecosystem monitoring in data-limited regions.