<p>The organic matter in kitchen waste can by anaerobically fermented to produce combustible biogas, which has broad prospects in alleviating the energy crisis. This study proposes a new method for high-precision prediction of biogas production based on small sample on-site sampling for the anaerobic digestion (AD) process of kitchen waste. Firstly, we proposed a novel feature selection method named MIC-mRMR to more efficiently select the optimal input feature subset from all the features collected in the field. Secondly, an improved small sample augmentation method (named SRG) is developed to address the issues of insufficient and unbalanced high-quality samples collected. Subsequently, we trained the gated recurrent unit (GRU) and built the high-precision prediction model for biogas production in AD process under imbalanced small sample conditions successfully. Finally, validation experiments and testing experiments are also conducted to further estimate the fitting and generalization performance of the established model. Compared with the other seven methods, the biogas production prediction model obtained by our proposed method has the highest accuracy with the R² (0.9903) closest to 1 and the smallest prediction errors. It can more effectively utilize the partial information collected during AD to make preliminary predictions on biogas production.</p>