Impact of neural network depth on split federated learning performance in low-resource UAV networks