Soil total nitrogen prediction using sentinel-2 simulated bands and machine learning: a laboratory spectroscopy study in Hemerocallis citrina Baroni fields

Soil total nitrogen (STN) is a crucial indicator of crop productivity and soil health. Accurate monitoring of STN is essential for optimizing nitrogen management and achieving sustainable agricultural development. An adequate and timely STN supply serves as a key physiological basis for promoting effective tillering, flower stalk development, and continuous multibatch bud formation in Hemerocallis citrina Baroni. To address the challenges posed by the high-dimensionality of hyperspectral data and the dynamic spectral response of STN across different growth stages, this study employed spectral resampling to select feature bands based on Sentinel-2 sensor data(Simulation of Sentinel-2 Bands, SSB method). Specifically, hyperspectral data were collected under laboratory controlled conditions (constant temperature darkroom, standard light source, air-dried ground soil), simulated Sentinel-2 sensor bands through spectral resampling (SSB method), and constructed an STN prediction framework based on 8 machine learning algorithms(random forest, extreme gradient boosting, back propagation neural network (BPNN), genetic algorithm-optimized BPNN (GA-BPNN), convolutional neural networks (CNN), and a hybrid CNN-bidirectional long short-term memory-attention model). The model performance was comprehensively evaluated using the coefficient of determination (R 2 ), root mean square error (RMSE), mean absolute error (MAE), and mean bias error (MBE). This study aims to establish laboratory-scale soil-spectral chemical relationship baselines, providing band selection and algorithm validation references for subsequent field remote sensing applications, rather than directly developing field operational systems. The results showed that: (1) the three-band spectral index TBI3 exhibited the highest correlation with STN across the full growth period (R=0.7354). The optimal indices for specific growth stages were TBI4, TBI3, and TBI5 for the spring seedling/leaf expansion, bolting/flowering, and bud emergence stages, respectively, with TBI-series indices exhibiting significantly superior performance compared to two-dimensional indices; (2) the GA-BPNN model achieved the highest accuracy for the full growth period, with a test R 2 of 0.6284, along with the lowest MAE (0.0693 g·kg -1 ) and RMSE (0.0879 g·kg -1 ), demonstrating outstanding generalization capability; and (3) the GA-BPNN model outperformed the other models in comparative analyses across different growth stages, and the growth stage-specific integrated modeling method showed higher prediction accuracy and enhanced resistance to overfitting (both training and test R 2 exceeded 0.6, with the gap reduced to 0.0064). Based on these findings, we propose a technical framework termed "SSB-SPXY-GA-BPNN-growth stage adaptation", which provides theoretical and methodological support for precise STN monitoring and variable-rate fertilization.