Using spectral information to detect chlorophyll content in rice canopy leaves quickly, non-destructively and accurately has a great practical significance for rice growth evaluation, precise fertilization and scientific management. In this paper, japonica rice in northeast China is taken as the research object, and rice canopy hyperspectral data of key growth stages are obtained through plot experiments. Firstly, the standard normal variate (SNV) is used to preprocess the spectral data, based on the processed spectral data and the random frog (RF) algorithm, by combining a correlation coefficient analysis method (CC) and the successive projections algorithm (SPA), an improved random frog algorithm (fpb-RF) is proposed, which combines two primary bands to select the feature bands of chlorophyll content, It is compared with the standard RF, CC and SPA methods, respectively. A hybrid prediction model (GPR-P) with gaussian process regression (GPR) compensation partial least squares regression (PLSR) is proposed: PLSR method is used to preliminarily predict the chlorophyll content in rice to obtain the linear trend of chlorophyll content, and then the GPR with good nonlinear approximation ability is used to predict the deviation of PLSR model, then the final prediction value is obtained by superposition of two outputs. To verify the superiority of the proposed method, with the feature bands by different extraction methods as inputs, PLSR, Least Square Support Vector Machine (LSSVM) and BP neural network prediction models are respectively established. The results show that under the same prediction model conditions, the improved fpb-RF algorithm can better reduce the complexity and improve the model’s prediction performance by extracting feature bands as input. Both the determination coefficient ($R^{2}_{p}$) of the test set and the determination coefficient ($R^{2}_{c}$) of each model’s training set are higher than 0.704 7. In addition, the $R^{2}_{c}$ and $R^{2}_{p}$ of the proposed GPR-P model are both higher than 0.755 3 when each algorithm extracts feature bands. Among them, the GPR-P model with the input of the feature band extracted by the fpb-RF method has the highest prediction accuracy, $R^{2}_{c}$ and $R^{2}_{p}$ are 0.781 5 and 0.779 6 respectively, RMSE-C and RMSE-P are 0.904 1 and 0.928 3 mg·L^-1 respectively, which provides a valuable reference for the detection and evaluation of chlorophyll content in northeast japonica rice.