Bangladesh, as an agro-based nation, is facing a double burden: growing population, increased food demand, and unprecedented use of fossil fertilizers. Artificial Neural Network (ANN) models were used in this study to predict agricultural methane and CO2 emissions using long-term agricultural data from Bangladesh (1972 to 2019). Different combinations of number of layers and neurons are used to choose the best regression ANN models for predicting agricultural methane and CO2 emissions (percentage of total), and the best model is chosen by trial and error using training SSE, Testing SSE, and RMSE as output measurement methods. This comparative study will provide a wonderful prediction ratio of methane and CO2 emission in agricultural fields and it will help the making policies or solutions for reducing methane and CO2 produced by agricultural fields. It will help the government of Bangladesh take steps and by taking, the solution will agri-production increase rapidly with minimal emission of methane and CO2. Our applied model holds on 95.33% accuracy for predicting agricultural methane and CO2 emissions in Bangladesh.