Rice is one of the most significant food crops all over the world. Biomass is a key phenotypic trait in rice research. A new method for nondestructive detection of rice biomass for multiple varieties at reproductive stage based on deep belief network was proposed. RGB images of 483 different rice varieties under normal growth environment and drought stress environment were captured at three time points: before stress, after stress and after rehydration. After image acquisition, the images were segmented by using fixed threshold in HSL color space and 57 image-derived features related to rice biomass were extracted. After data normalization, a rice biomass model was built based on deep belief network. The influences of visible layer type, hidden layer number, hidden layer neuron number, learning rate, epoch number and momentum on the performance of deep belief network were tested. The best model was selected based on the coefficient of determination (R2), mean absolute percent error (MAPE) and standard deviation of absolute percent error (SAPE). The deep belief network model was also compared with the stepwise linear regression model. The results showed that the biomass measurement model based on the deep belief network performed better (R2 was 0.9299, MAPE was 11.19% and SAPE was 18.36%). The research offered a new nondestructive method for accurately measuring rice biomass for multiple varieties under different growth environments, which would provide a new tool for rice research.