基于MAML改进方法的遥感图像分类方法

Remote Sensing Image Classification Method Based on MAML Improvement

摘要: 遥感图像分类是遥感技术中的核心任务，旨在根据图像的光谱、空间和纹理信息对地表物体进行分类。尽管深度学习方法在遥感图像分类中取得了显著进展，但大规模标注数据的需求仍然是一个挑战。为解决这一问题，元学习(Meta-Learning)作为一种有效的小样本学习技术，近年来在遥感图像分类中受到了广泛关注，特别是Model-Agnostic Meta-Learning (MAML)方法。然而，MAML在遥感图像分类中的应用面临跨域迁移和类别不平衡等问题。文章提出了一种基于改进MAML的遥感图像分类方法，旨在提高少样本条件下的分类精度，并解决跨域迁移和类别不平衡问题。具体而言，文章结合扩散模型(Diffusion Model)进行数据增强，增加样本数量，改善数据分布，从而提高模型的鲁棒性和泛化能力。同时，通过改进MAML的梯度更新策略，结合导数顺序退火(DA)方法，使模型在不同阶段采用不同阶的导数进行更新，增强了模型的适应性和稳定性。实验结果表明，该方法在UC Merced Land-Use、NWPU-RESISC45和Mini-ImageNet数据集上的分类精度优于传统方法。在UC Merced数据集上，分类精度达到98.06%，在NWPU-RESISC45数据集上达到95.76%，在Mini-ImageNet数据集上也取得了良好的分类效果，验证了其不仅在遥感图像分类中具有有效性和优势，还具有较强的泛用性。

Abstract: Remote sensing image classification is a core task in remote sensing technology, aiming to classify surface objects based on spectral, spatial, and textural information of the images. Although deep learning methods have made significant progress in remote sensing image classification, the need for large-scale labelled data remains a challenge. To solve this problem, Meta-Learning (MAML), as an effective small-sample learning technique, has received much attention in remote sensing image classification in recent years, especially the Model-Agnostic Meta-Learning (MAML) method. However, the application of MAML in remote sensing image classification faces problems such as cross-domain migration and category imbalance. In this paper, we propose a remote sensing image classification method based on improved MAML, which aims to improve the classification accuracy under fewer sample conditions and solve the problems of cross-domain migration and category imbalance. Specifically, this paper combines the Diffusion Model (DM) for data enhancement to increase the number of samples and improve the data distribution so as to improve the robustness and generalization ability of the model. Meanwhile, by improving the gradient update strategy of MAML, combined with the derivative order annealing (DA) method, the model is updated with different orders of derivatives at different stages, which enhances the adaptability and stability of the model. The experimental results show that the classification accuracy of this paper’s method on UC Merced Land-Use, NWPU-RESISC45, and Mini-ImageNet datasets outperforms that of the traditional method. The classification accuracy reaches 98.06% on the UC Merced dataset, 95.76% on the NWPU-RESISC45 dataset, and also achieves good classification results on the Mini-ImageNet dataset, which verifies that the method is not only effective and advantageous but also highly generalizable in remote sensing image classification.