Sentinel-2-Based Estimation and Mapping of Rice Grain Total Amino Acid Content Using Fractional-Order Multiband Spectral Indices and Feature Selection

This is the authors' abstract. We don't add key points for this paper.

Reliable satellite-based estimation of rice grain total amino acid content (TAA) remains challenging because TAA is an indirect grain-quality trait with weak expression in broad-band canopy spectra, while multiband feature construction generates a high-dimensional predictor space relative to the limited number of field samples. This study therefore examined whether TAA-related multispectral information could be effectively enhanced, retained, and predicted under this small-sample, high-dimensional setting. Sentinel-2 Level-1C imagery and 31 composite rice grain samples collected in the Daying River Valley, Yingjiang County, China, in 2019 were used to develop a framework based on fractional-order derivative (FOD) transformation, multiband spectral-index construction, feature selection, and machine learning prediction. FOD orders of 1.2–1.8 generally strengthened spectral associations with TAA, and three-band indices showed stronger associations than two-band indices, with TBI1 (B12, B3, B7) reaching the highest correlation at order 1.6 (r = 0.641). Ablation analysis further showed that predictive performance improved progressively as FOD-transformed and multiband features were incorporated beyond the original Sentinel-2 bands and conventional vegetation indices. Elastic Net + SVM achieved the highest stratified out-of-fold performance (R2 = 0.886, RMSE = 0.264%, RPD = 2.96), whereas Boruta + Stacking performed best under both four- and five-block spatial cross-validation (R2 = 0.7489 and 0.7750, respectively). These results indicate that FOD-based multiband feature construction can recover useful information associated with rice grain TAA from broad-band Sentinel-2 observations, while the reduction in performance under spatial separation highlights the need for the cautious interpretation of within-study cross-validation accuracy. The resulting maps should therefore be regarded as model-assisted estimates of within-region TAA variation that can support candidate quality-zone identification and targeted field sampling under the sampled 2019 Yingjiang conditions. Validation across additional years, cultivars, and regions is required before broader operational application.