随着遥感卫星技术的发展，遥感图像的获取变得容易而且具有高分辨率，这些图像是遥感领域各种应用的重要数据来源。遥感图像场景分类是遥感图像应用领域的基本技术。遥感图像场景分类技术在智慧农业、灾害预警和智慧城市规划等领域得到了应用。但是在处理遥感图像场景分类时通常会面对数据不平衡、有标签样本少和无标签样本利用率低等问题，在实际应用中也面临准确高效等问题。针对以上问题，本文进行了如下四部分的工作：（1）构建了贵州遥感图像数据集，包含不平衡场景的数据分布和一般场景下的数据分布。该数据集包含了12个日常场景类别的标签数据4820张和无标签数据41790张，为遥感图像场景分类的研究和应用提供数据基础。（2）提出了一种基于自监督对比学习的遥感图像场景分类算法。本文采用了基于动量更新孪生网络的自监督对比学习策略，在预训练的过程中解决数据不平衡和无标签数据利用率低的情况。在下游任务中，本文采用迁移学习解决有标签样本少的情况，并使用权重代价策略进行训练缓解数据不平衡情况。实验结果表明，该方法能缓解遥感图像分类面临的问题，提高了不平衡场景下的遥感图像分类效果。（3）提出了一种基于周期自然指数退火的多视图集成方法。在训练层面上，本文提出了周期自然指数退火的学习率衰减方式，在一次训练过程中保存模型的多个权重用于集成学习。在数据层面上，本文将原始数据进行几何变换和灰度变换生成多个视图，从而利用不同视图的数据训练模型并集成。本文方法将周期自然指数退火集成方法和多视图集成方法结合，在一个模型架构上实现了二阶集成学习，满足了工业应用高效准确的需求。实验表明，该方法能够提高一般场景下遥感图像分类效果。（4）构建了一种遥感图像识别系统。本文依托于研究内容的需求和技术框架，建立了遥感图像场景分类识别系统，辅助业务人员，帮助企业实现遥感图像的数字化管理和智能化流程。

With the development of remote sensing satellite technology, remote sensing images are becoming clearer and easier to obtain. These images are important data sources for various applications in the field of remote sensing. Remote sensing image scene classification is a basic technology in the field of remote sensing image applications. This technology has been applied in the fields of smart agriculture , disaster warning and smart city planning. When processing remote sensing image scene classification, problems such as data imbalance, fewer labeled samples, and low utilization of unlabeled samples are usually encountered. In practical applications, there are issues of accuracy and efficiency. In response to the above issues, this article carried out the following four parts of work:(1) A Guizhou remote sensing image dataset was constructed, including data distribution for unbalanced scenes and general scenes. The dataset contains 4820 labeled data of 12 daily scene categories and 41790 unlabeled data, which provides a data foundation for the research of remote sensing image scene classification. (2) A remote sensing image scene classification algorithm based on self supervised contrastive learning is proposed. In this paper, A self supervised contrastive learning strategy based on momentum update siamese networks is adopted to solve the situation of data imbalance and low utilization of unlabeled data during the pre-training process. In the downstream tasks, transfer learning is used to solve the situation of few labeled samples, and the weight cost strategy is used for training to alleviate the data imbalance. Experimental results show that this method can alleviate the problems faced by remote sensing image classification and improve the remote sensing image classification effect in unbalanced scenes. (3) A multi view integration method based on periodic natural exponential annealing is proposed. At the training level, this paper proposes a learning rate decay method of periodic natural exponential annealing, which saves multiple weights of the model for ensemble learning during a training process. At the data level, this paper performs geometric transformation and grayscale transformation on the original data to generate multiple views, so as to use the data of different views to train the different models. The method in this paper combines the periodic natural exponential annealing ensemble method with the multi-view ensemble method to achieve second-order ensemble learning on a model architecture. This second-order ensemble method meets the efficient and accurate requirements of industrial applications. Experiments show that this method can improve the classification effect of remote sensing images in general scenes. (4) A remote sensing image recognition system. Based on the requirements of the research content and the technical framework, this article established a remote sensing image scene recognition system to assist business personnel and help enterprises achieve digital management and intelligent processes of remote sensing images.