多尺度卷积神经网络的遥感影像多标签分类

Multi-label Classification of Remote Sensing Image Based on Multi-scale Convolutional Neural Network

卫星、飞机等对地观测系统迅速发展,大量的遥感影像含有多种地物但不能得到有效利用,针对这一问题,本文提出了采用一种多尺度卷积神经网络对遥感影像进行多标签分类,以帮助影像的管理和理解。该算法通过获取更多语义信息和标签特征来提高分类精度。实验基于DLRSD数据集与XGBoost算法和基础网络进行了比较,且通过4个对比实验证明该算法有效可行,并对这些对比实验的内存及训练时间进行了统计。相比于XGBoost算法,本研究方法 F2分数高了0.062;相比于基础网络,本研究方法 F2分数高了0.088,其中,4个改进分别将F2分数提高了0.008、0.017、0.015、0.048;内存及训练时间对比中,最后一个改进使训练时间增加到两倍左右,其他改进对内存和训练时间改动不大。本文提出的多尺度卷积神经网络方法,虽然训练时间增加但使遥感影像多标签分类的精度提升很大,具有可操作性。

Satellite and aircraft ground observation systems have developed rapidly. A large number of remote sensing images contain many features but cannot be effectively utilized. To solve this problem,a multi-scale convolutional neural network is proposed to classify remote sensing images by multi-label. Help image management and understanding. The algorithm improves classification accuracy by obtaining more semantic information and label features. The experiment is based on the DLRSD data set compared with the XGBoost algorithm and the underlying network,and the improvement is valid and feasible through four comparative experiments,and the memory and training time of these comparative experiments are statistically analyzed. Compared with the XGBoost algorithm,the F2 score of this research method is higher by 0.062;compared with the basic network,the F2 score of this research method is higher by 0.088,and four of the improvements improve the F2 score by 0.008,0.017,0.015,and 0.048 respectively;In the training time comparison,the last improvement increased the training time by a factor of two. Other improvements did not change the memory and training time.The proposed multi-scale convolutional neural network method,although the training time is increased,makes the accuracy of multilabel classification of remote sensing images greatly improved and operable.