Title: Enhancing Leaf Spectral Reflectance Forecasting via Multi-Head Attention Neural Networks

Abstract:

Precise simulation of leaf spectral reflectance based on physiological and biochemical characteristics is a critical step toward advancing remote sensing technologies in precision agriculture and plant science. While radiative transfer models like PROSPECT-PRO are widely adopted, they depend on generalized trait-reflectance correlations derived from diverse species. Consequently, these models may fail to fully represent the unique spectral dynamics of particular crops, such as grapevines. To address this limitation, we engineered a data-driven model that predicts spectra from traits using a multi-head attention neural network. This architecture was trained on a specialized grapevine dataset comprising 16 distinct leaf traits, observed across various varieties, developmental stages, and years.

The model’s performance was assessed through stratified 5-fold cross-validation, yielding an average coefficient of determination (R²) of 0.84 and a normalized root mean squared error (NRMSE) of 1.52 percent. These metrics indicate strong accuracy and robust generalizability. Furthermore, when benchmarked against PROSPECT-PRO in forward mode, the neural network demonstrated a reduced mean absolute error (MAE), particularly within the near-infrared (NIR) and shortwave-infrared (SWIR) spectral bands. These findings underscore the value of species-specific modeling strategies and highlight how incorporating both structural and biochemical traits into data-driven frameworks can markedly enhance spectral prediction capabilities. Ultimately, this proposed framework offers a reliable method for producing precise leaf-level reflectance data, supporting applications in canopy trait retrieval, vineyard surveillance, and remote sensing-based crop management.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC