Federated learning is an effective method of solving the challenge of using marine ecosystem image data to detect marine plastic pollution since different datasets are scattered across numerous sources. The current research presents the Federated Evolutionary Stochastic Gradient Descent Learning (FedEvoSGD) method, an innovative framework for the effective detection of marine plastic pollution in federated settings. By integrating mutual conditional learning and knowledge distillation with federated learning, FedEvoSGD offers a novel method for model training and calibration. Bidirectional learning enhances the model’s accuracy by allowing dynamic adjustments to local dataset variances. Through testing various hyperparameter tuning algorithms, this paper makes an important contribution by showing that GridSearch and RandomSearch are the best methods for federated learning fine-tuning. In addition, this research contrasts and compares denoising approaches, cross-validation approaches, and data augmentation approaches to offer insights improving the proposed model’s flexibility. These slight modifications, coupled with the FedEvoSGD algorithm, enhance our comprehension and capacity for utilization of federated learning to find marine plastic debris. With the use of a hybrid learning paradigm, thorough parameter exploration, and demanding validation procedures, the research lays the groundwork for the development of reliable and efficient marine plastic pollution detection systems in collaborative environments. The level of accuracy obtained by the ultimate proposed model FedEvoSGD+GridSearch provided a training accuracy value of 81.78%.