In order to explore a better hyperspectral inversion model for monitoring nitrogen content in rice canopy leaves by remote sensing, based on rice plot experiments, the canopy height spectral data of rice at different growth stages were obtained. Based on the comprehensive comparison of the first derivative (1-Der), standard normal variable transformation (SNV) and SG smoothing method, a spectral processing method (SNV-FDSGF) combining standard normal variable transformation with SG filtering method of first derivative was proposed. The sensitive bands of different growth stages were screened out by non-information variable - competitive adaptive reweighted sampling method (UVE-CARS). Two sensitive bands of each growth period were randomly combined to construct a difference spectrum index DSI (difference spectral index), a ratio spectral index RSI (ratio vegetation index) and a normalized spectrum index NDSI (normalized defference spectral index) with high correlation with nitrogen content in rice leaves. Among them, the optimal vegetation index and determination coefficient R2 at the tillering, jointing and heading stages were: DSI(R857, R623), 0.704; DSI(R670, R578), 0.786; DSI(R995, R508), 0.754. Using the superior three planting indices in each growth period as inputs, the adaptive differential optimization extreme learning machine (SaDE-ELM), radial basis function (RBF-NN) and particle swarm optimization BP neural network (PSO-BPNN) inversion models were constructed respectively. The results showed that SaDE-ELM had the best modeling effect. Compared with RBF-NN and PSO-BPNN, the stability and prediction ability of the model were significantly improved. The determination coefficient R2 of training set and verification set of each growth phase inversion model was above 0.810 and RMSE was below 0.400, which could provide certain theoretical basis for quantitative prediction of nitrogen content in rice canopy leaves.