Integration of Super-Resolution Remote Sensing and Feature Optimization for Accurate and Interpretable Nitrogen Balance Index Estimation in Winter Wheat

Abstract

The Nitrogen Balance Index (NBI) is a key indicator for assessing the nitrogen status of crops; however, conventional monitoring techniques are often invasive, expensive, and challenging to implement over large areas. This study focuses on early-season winter wheat in the Guanzhong region and integrates super-resolution reconstruction technology (Sen2Res) with Sentinel-2 imagery and a multi-feature combination optimization approach to develop a high-accuracy machine learning model for NBI estimation. Additionally, SHAP interpretability is employed to elucidate the underlying mechanisms of the model. The results demonstrated that Sen2Res super-resolution imagery captured spatial details and textures with higher precision. Furthermore, the reflectance values, spatial texture parameters, and vegetation indices extracted from these images showed stronger correlations with NBI compared to those from the original imagery. Using a two-stage feature selection process, which included the correlation coefficient method followed by recursive feature elimination, eight important variables were identified. These mainly included the red-edge band B7 (740 nm), near-infrared band B8A (865 nm), and the vegetation index CI740-B8A. Among the tested models, the random forest trained with the optimized feature set (RR-RFE-DRFR) achieved superior performance, with an R2 of 0.77 and an RMSE of 1.57 on the test set—an improvement of approximately 20% in accuracy over traditional linear models. SHAP analysis indicated that red-edge bands and near-infrared vegetation indices accounted for up to 75% of the contribution to NBI prediction, substantiating the relevance of spectral response mechanisms associated with leaf nitrogen metabolism. This study presents improvements in three main areas: (1) the integration of Sen2Res super-resolution reconstruction with multi-source feature fusion, which overcomes the limitations of fine-scale farmland monitoring due to the insufficient resolution of original imagery; (2) the development of a two-stage feature optimization strategy that combines correlation analysis (RR: correlation coefficient) and recursive feature elimination (RFE) to reduce the loss of model generalization caused by high-dimensional remote sensing data redundancy; and (3) the incorporation of the SHAP interpretability framework into remote sensing-based estimation of crop nitrogen status, thereby enhancing the transparency of machine learning “black-box” models and providing a theoretical foundation for designing variable-rate fertilization prescriptions. The proposed technical framework—comprising super-resolution imagery, multi-source feature optimization, and interpretable modeling—enables regional-scale monitoring of NBI in winter wheat and supports precision fertilization decision-making.

Keywords: Nitrogen Balance Index, Sentinel-2, Super-Resolution Reconstruction, machine learning, SHAP Interpretability