| 智能优化学习的高空间分辨率遥感影像语义分割 |
| |
| 引用本文： | 邵振峰,孙悦鸣,席江波,李岩.智能优化学习的高空间分辨率遥感影像语义分割[J].武汉大学学报(信息科学版),2022,47(2):234-241. |
| |
| 作者姓名： | 邵振峰 孙悦鸣 席江波 李岩 |
| |
| 作者单位： | 1.武汉大学测绘遥感信息工程国家重点实验室，湖北 武汉，430079 |
| |
| 基金项目： | 国家自然科学基金41771454国家自然科学基金61806022江西省03专项及5G项目20212ABC03A09内蒙古自治区自然科学基金2019MS04017内蒙古自治区高等学校科学研究项目NJZY20277地理信息工程国家重点实验室科研项目SKLGIE2018-M-3-4长安大学中央高校基本科研业务费项目300102269103长安大学中央高校基本科研业务费项目300102269304长安大学中央高校基本科研业务费项目300102269205 |
| |
| 摘 要： | 高空间分辨率遥感影像正被广泛应用，而传统分类算法在高分遥感影像上的精度和效率较差，深度学习语义分割算法在实际分类中泛化性较差。为了适应大范围高分遥感影像的特点，提出了一种基于U-Net网络的模拟退火超参数优化与深度可分离卷积语义分割模型。首先在U-Net网络基础上使用了深度可分离卷积模块来进行特征提取，在保持高效性的同时减少模型的参数量和计算量，然后利用基于模拟退火的智能优化学习模型搜索网络超参数的全局最优解，自动优化网络训练初始点，最后在ISPRS2D和GID(Gaofen image dataset)数据集上进行实验。对比实验结果表明，在ISPRS2D数据集的分类结果中，建筑物、低植被和汽车及总体分类精度均有提高，在GID数据集的分类结果中，水域、草地、森林及总体分类精度均有大幅提高。实验结果验证了所提模型的高效性、高精度性与鲁棒性。
|
| 关 键 词： | 语义分割 高分辨率遥感影像 深度学习 超参数优化 |
| 收稿时间： | 2020-12-16 |
Intelligent Optimization Learning for Semantic Segmentation of High Spatial Resolution Remote Sensing Images |
| |
| Institution: | 1.State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China2.College of Geology Engineering and Surveying, Chang'an University, Xi'an 710054, China3.Inner Mongolia Electronic Information Vocational Technical College, Huhhot 010070, China |
| |
| Abstract: | Objectives With the hardware development of remote sensing sensors, the high spatial resolution remote sensing images are widely used. In the actual classification, typical classification algorithms can not achieve high accuracy and efficiency, while deep learning semantic segmentation algorithms can not achieve good generalization. Methods In order to adapt to the large-scale high-resolution images, we design a semantic segmentation model of simulated annealing hyperparameter optimization and depthwise separable convolution based on U-Net. Firstly, the depthwise separable convolutional module is used to extract the features on the baseline. Then, the intelligent optimization learning model based on simulated annealing searches the global optimal solution of hyperparameters. Finally, experiments were carried out in ISPRS2D and Gaofen image dataset (GID). Results Compared to other classification methods, the proposed method achieves the highest classification accuracy in buildings, low vegetation, trees, cars and overall accuracy, and the total accuracy of classification results reaches 86.5% in ISPRS2D. And in the classification results of GID, the accuracies of the proposed method in water, grassland, forest and general classification are all greatly improved. The mean intersection over union of the proposed method is also higher than that of other methods. Conclusions The results demonstrated the higher efficiency, higher accuracy and robustness of the proposed method. |
| |
| Keywords: | |
|
| 点击此处可从《武汉大学学报(信息科学版)》浏览原始摘要信息 |
| 点击此处可从《武汉大学学报(信息科学版)》下载免费的PDF全文 |
|
