Federated learning often experiences slow and unstable convergence due to edge-side data heterogeneity. This problem becomes more severe when edge participation rate is low, as the information collected from different edge devices varies significantly. As a result, communication overhead increases, which further slows down the convergence process. To address this challenge, we propose a simple yet effective federated learning framework that improves consistency among edge devices. The core idea is clusters the lookahead gradients collected from edge devices on the cloud server to obtain personalized momentum for steering local updates. In parallel, a global momentum is applied during model aggregation, enabling faster convergence while preserving personalization. This strategy enables efficient propagation of the estimated global update direction to all participating edge devices and maintains alignment in local training, without introducing extra memory or communication overhead. We conduct extensive experiments on benchmark datasets such as Cifar100 and Tiny-ImageNet. The results confirm the effectiveness of our framework. On CIFAR-100, our method reaches 55% accuracy with 37 fewer rounds and achieves a competitive final accuracy of 65.46%. Even under extreme non-IID scenarios, it delivers significant improvements in both accuracy and communication efficiency. The implementation is publicly available at https://github.com/sjmp525/CollaborativeComputing/tree/FedCCM (accessed on 20 October 2025).

Semantic segmentation provides important technical support for Land cover/land use (LCLU) research. By calculating the cosine similarity between feature vectors, transformer-based models can effectively capture the global information of high-resolution remote sensing images. However, the diversity of detailed and edge features within the same class of ground objects in high-resolution remote sensing images leads to a dispersed embedding distribution. The dispersed feature distribution enlarges feature vector angles and reduces cosine similarity, weakening the attention mechanism’s ability to identify the same class of ground objects. To address this challenge, remote sensing image information granulation transformer for semantic segmentation is proposed. The model employs adaptive granulation to extract common semantic features among objects of the same class, constructing an information granule to replace the detailed feature representation of these objects. Then, the Laplacian operator of the information granule is applied to extract the edge features of the object as represented by the information granule. In the experiments, the proposed model was validated on the Beijing Land-Use (BLU), Gaofen Image Dataset (GID), and Potsdam Dataset (PD). In particular, the model achieves 88.81% for mOA, 82.64% for mF1, and 71.50% for mIoU metrics on the GID dataset. Experimental results show that the model effectively handles high-resolution remote sensing images. Our code is available at https://github.com/sjmp525/RSIGT (accessed on 16 April 2025).