滑坡灾害是造成人员伤亡和经济损失最严重的地质灾害之一，利用遥感影像快速、准确地检测滑坡，确定滑坡位置、范围等信息对灾后救援具有重要的意义。同时，遥感技术的迅速发展使得海量高分辨率遥感数据传入地面，传统的滑坡检测方法难以满足海量遥感影像滑坡高精度实时检测的需求。深度学习方法具有自学特征的能力和较强的非线性特征表达能力，近年来在滑坡检测领域逐渐被应用，但此类研究仍然不够充分。

本文从数据增强和方法改进两方面出发，以目前工业界检测性能表现优异的YOLOv5算法为基础，针对滑坡样本量不足导致的模型泛化能力差和滑坡检测精度不高等问题开展研究，具体内容如下：

（1）为解决基于高分辨率遥感影像的滑坡样本数量有限，滑坡样本多样性差的问题，本文在已获取样本的基础上分别进行了像素级样本数据增强、图片级样本数据增强、困难样本构建、小目标样本重组、边缘样本构建等数据增强操作。通过多种方式的数据增强提高滑坡样本的多样性和丰富性，从而提高滑坡检测模型的泛化能力和检测精度。

（2）基于卷积神经网络（Convolutional Neural Networks, CNNs）模型构建原理分析了融合多种主干网络的YOLOv5模型滑坡特征提取过程。对卷积神经网络基本组成单元，包括卷积层、池化层、归一化层、激活函数等模块的滑坡特征提取特点进行了可视化分析。然后对8种主流特征提取网络进行了分析，并将这8种网络与YOLOv5模型融合进行了滑坡检测实验，最后对融合后各种模型的滑坡检测性能进行评价。

（3）将非局部信息统计操作算子与CNNs结合，在CNNs中融合具有全局信息获取能力的非局部信息统计注意力机制（No-Local Attention, NLA），以YOLOv5模型为基础，构建了一种非局部信息统计注意力的YOLOv5滑坡检测模型（NLA-YOLOv5）。首先，将No-Local Attention引入主干网络中提高模型获取全局信息的能力；然后，改进了一种大卷积核空间金字塔池化模块（Large Kernel Spatial Pyramid Pooling, LK-SPP）代替普通卷积空间金字塔池化模块（Spatial Pyramid Pooling Fsat, SPPF）以提高网络的有效感受野；最后，通过调整LK-SPP中卷积核大小组合确定NLA-YOLOv5框架。实验结果表明：NLA-YOLOv5滑坡检测精度提升到80%，相较于改进前提升了5%，精确度、召回率和F1值分别提升了7%、2%和5%。模型检测速度达到69f/s，在保证模型实时检测的前提下，有效提高了滑坡检测能力。

（4）针对滑坡样本获取难度大，样本数据数量少，训练模型泛化能力差的问题。本文利用深度卷积生成对抗网络（Deep Convolutional Generative Adversarial Networks, DCGAN）进行滑坡样本数据增强。以渝东南地区为研究区，基于谷歌地球三通道光学遥感影像获取滑坡训练样本，训练生成对抗网络模型，并利用训练好的模型进行滑坡样本生成，最终生成滑坡样本3429个。利用增强后的滑坡样本数据集对NLA-YOLOv5进行训练，并测试滑坡检测精度，结果表明，基于DCGAN的滑坡数据增强方式能有有效提高滑坡检测模型的泛化能力。最后，在公开的毕节市滑坡样本数据集上验证了NLA-YOLOv5模型的有效性和在不同数据集上的可迁移性。

Landslide disasters are one of the most severe geological hazards that cause casualties and economic losses. It is of great significance to use remote sensing images to quickly and accurately detect landslides, determine the location, range, and other information for post-disaster rescue. Meanwhile, the rapid development of remote sensing technology has led to the massive influx of high-resolution remote sensing data on the ground, making traditional landslide detection methods insufficient to meet the high-precision real-time detection needs of massive remote sensing images of landslides. Deep learning methods have the ability to learn features autonomously and express strong non-linear features, and have gradually been applied in the field of landslide detection in recent years. However, such research is still not sufficient.

This article focuses on data augmentation and method improvement to address the issues of poor model generalization and low detection accuracy in landslide detection due to insufficient sample sizes. Based on the YOLOv5 algorithm, which has excellent performance in industrial detection, the research in this article is divided into two parts. The specific contents are as follows:

(1)In order to solve the problem of limited landslide samples and poor sample diversity based on high-resolution remote sensing images, this paper carried out pixel-level sample data augmentation, image-level sample data augmentation, difficult sample construction, small target sample reorganization, edge sample construction and other data augmentation operations based on the obtained samples. Through various data augmentation methods, the diversity and richness of landslide samples were improved, thereby improving the generalization ability and detection accuracy of the landslide detection model.

(2)Based on the construction principle of Convolutional Neural Networks (CNNs) models, this paper analyzed the landslide feature extraction process of the YOLOv5 model that integrates multiple backbone networks. The landslide feature extraction characteristics of the basic components of the convolutional neural network, including the convolution layer, pooling layer, normalization layer, activation function, and other modules were visualized and analyzed. Then, eight mainstream feature extraction networks were analyzed, and these eight networks were fused with the YOLOv5 model for landslide detection experiments. Finally, the landslide detection performance of the various models fused was evaluated.

(3)By combining non-local information statistical operator with CNNs and integrating the non-local information statistical attention mechanism (No-Local Attention, NLA) with global information acquisition ability into the CNNs, a non-local information statistical attention-based YOLOv5 landslide detection model (NLA-YOLOv5) was built on the basis of the YOLOv5 model. Firstly, No-Local Attention was introduced into the backbone network to improve the model's ability to acquire global information. Then, a large kernel spatial pyramid pooling module (LK-SPP) was improved to replace the common convolutional spatial pyramid pooling module (Spatial Pyramid Pooling Fast, SPPF) to enhance the effective receptive field of the network. Finally, the NLA-YOLOv5 framework was determined by adjusting the convolution kernel size combination in the LK-SPP. The experimental results show that the accuracy of NLA-YOLOv5 landslide detection has been improved to 80%, an increase of 5% compared to the previous improvement, and the precision, recall, and F1 values have been improved by 7%, 2%, and 5% respectively. The detection speed of the network model reached 69f/s, effectively improving the landslide detection ability while ensuring real-time detection of the model.

(4)In view of the difficulty in obtaining landslide samples, the number of sample data is small, and the generalization ability of the training model is poor, this paper employs Deep Convolutional Generative Adversarial Networks (DCGAN) for landslide sample data augmentation. Using three-channel optical remote sensing images from Google Earth, we obtained landslide training samples from the southeastern Chongqing area, trained the generative adversarial network model, and generated 3429 landslide samples. The enhanced landslide dataset was used to train NLA-YOLOv5, and the landslide detection accuracy was tested. The results show that DCGAN-based landslide data augmentation improves the generalization ability of the landslide detection model effectively. Finally, the effectiveness and transferability of the NLA-YOLOv5 model on different datasets were verified using a public landslide sample dataset.