基于集成卷积神经网络的遥感影像场景分类 下载: 1289次
Scene Classification of Remote Sensing Images Based on Integrated Convolutional Neural Networks
1 中国科学院长春光学精密机械与物理研究所, 吉林 长春 130033
2 中国科学院大学, 北京 100049
3 长光卫星技术有限公司吉林省卫星遥感应用技术重点实验室, 吉林 长春 130102
图 & 表
图 1. 集成神经网络结构
Fig. 1. Architecture of integrated neural network
下载图片 查看原文
图 2. 集成网络构建流程
Fig. 2. Flow chart of integrated network construction
下载图片 查看原文
图 3. NWPU-RESISC45数据集的场景图
Fig. 3. Scene images of NWPU-RESISC45 dataset
下载图片 查看原文
图 4. ResNet-50训练过程中准确率、损失值和学习率随循环次数的变化。(a)准确率;(b)损失值;(c)学习率
Fig. 4. Accuracy, loss value and learning rate versus number of cycles in training process of ResNet-50.(a) Accuracy; (b) loss value; (c) learning rate
下载图片 查看原文
图 5. CNN的分类结果
Fig. 5. Classification results based on CNN
下载图片 查看原文
图 6. BP网络的训练过程中准确率和损失值随循环次数的变化曲线。(a)准确率;(b)损失值
Fig. 6. Accuracy and loss value versus number of cycles in training process of BP network. (a) Accuracy; (b) loss value
下载图片 查看原文
图 7. 集成模型对数据集进行分类预测后得到的混淆矩阵
Fig. 7. Confusion matrix obtained after classification prediction of dataset by integrated model
下载图片 查看原文
图 8. 与其他算法单类的准确率对比
Fig. 8. Single accuracy comparison with those of other algorithms
下载图片 查看原文
图 9. 多种模型的分类准确率和预测时间
Fig. 9. Classification accuracies and prediction time of various models
下载图片 查看原文
图 10. 使用AlexNet进行分类的场景类别数量对集成模型的性能影响
Fig. 10. Impact of number of scene categories classified by AlexNet on performance of integrated model
下载图片 查看原文
表 1各个网络的训练参数及结果
Table1. Training parameters and results of each network
Model | Input size /(pixel×pixel) | Batch size /frame | Number of cycles | Training accuracy /% |
---|
Experiment I | Experiment II |
---|
AlexNet | 224×224 | 256 | 300 | 81.22 | 85.46 | ResNet-50 | 224×224 | 256 | 300 | 86.52 | 90.52 | ResNet-152 | 224×224 | 128 | 600 | 85.11 | 90.11 | DenseNet-169 | 224×224 | 128 | 600 | 82.44 | 87.44 | VGG-16[2] | - | - | - | 87.15 | 90.36 | Proposed model | - | - | - | 88.47 | 92.53 |
|
查看原文
表 2几种算法的性能对比
Table2. Performance comparison among several algorithms
Method | Color-histogram | BoVW | VGG-16 | ResNet-50 | Proposed | Competition |
---|
Accuracy /% | 27.52 | 44.97 | 90.36 | 90.59 | 92.53 | 93.41 | Standard deviation | 0.2184 | 0.2051 | 0.0673 | 0.0657 | 0.0593 | 0.0451 | Prediction time /s | - | - | 0.62 | 0.47 | 0.41 | 2.26 |
|
查看原文
表 3与其他方法的平均准确率对比
Table3. Average accuracy comparison with those of other algorithms
Method | Accuracy /%(experiment I) | Accuracy /%(experiment II) |
---|
GIST[2] | 15.90 | 17.88 | LBP[2] | 19.20 | 21.74 | Color histograms[2] | 24.84 | 27.52 | BoVW+SPM[2] | 27.83 | 32.96 | LLC[2] | 38.81 | 40.03 | BoVW[2] | 41.72 | 44.97 | GoogLeNet[2] | 82.57 | 86.02 | VGG-16[2] | 87.15 | 90.36 | AlexNet[2] | 81.22 | 85.16 | Two-streamDFF[13] | 80.22 | 83.16 | ResNet-50 | 87.69 | 90.59 | Proposed model | 89.34 | 92.53 |
|
查看原文
张晓男, 钟兴, 朱瑞飞, 高放, 张作省, 鲍松泽, 李竺强. 基于集成卷积神经网络的遥感影像场景分类[J]. 光学学报, 2018, 38(11): 1128001. Xiaonan Zhang, Xing Zhong, Ruifei Zhu, Fang Gao, Zuoxing Zhang, Songze Bao, Zhuqiang Li. Scene Classification of Remote Sensing Images Based on Integrated Convolutional Neural Networks[J]. Acta Optica Sinica, 2018, 38(11): 1128001.