An integrated learning model integrating random forest and deep neural network is constructed to optimize the green financial resource allocation path in agricultural supply chain. The model introduces feature crossing mechanism and risk bias adjustment structure based on the traditional multi-task regression framework to improve the prediction accuracy of carbon emission intensity and credit matching. The experiment is based on 14,286 sets of agricultural green development data for comparative analysis, and the results show that the model reduces the green credit matching error by 43.1% compared with the support vector machine, and reduces the carbon emission score fitting error by 22.3% compared with the gradient boosting tree, with a comprehensive classification accuracy of 91.7%. The model has good generalization ability and deployment value.