基于相容粗糙集的BP神经网络湿地覆被信息提取——以双台子河口湿地为例

BP神经网络因具有自学习、自适应、大规模并行处理等特点而广泛应用于遥感影像分类中,但是该方法训练时容易陷入局部极小值,且收敛速度较慢,针对这些不足提出一种基于相容粗糙集的BP神经网络分类方法。本文以双台子河口湿地为研究对象,以Landsat-8 OLI影像为数据基础,利用相容粗糙集理论对样本数据集进行预处理,将得到的数据作为新的训练样本,在Matlab软件平台下建立BP神经网络的湿地覆被分类模型,进行湿地覆被信息提取,将分类结果与单纯的BP神经网络以及粗糙集样本属性约简预处理的分类结果进行比较分析。结果表明,基于相容粗糙集的BP神经网络分类方法可以剔除训练样本中的噪声数据,提高网络的训练成功率,缩短网络的收敛时间,分类效果较好,其总体精度达到91.25%,Kappa系数为0.8969,比单纯的BP神经网络分类结果高7.92%和0.0926,比粗糙集样本属性约简预处理方法的分类结果高3.03%和0.0357,是一种有效的湿地覆被分类方法。

Wetland land-cover information extraction of BP neural network based on tolerant rough set in Shuangtaizi estuarine wetland

Remote sensing data is the main resource for wetland monitoring because of its rich information. The BP neural network is widely used in remote sensing image classification with the characteristics of self-learning,adaptive and massively parallel processing. However,it is easy to fall into local minimum value,with slow convergence speed. A BP neural network classification method based on tolerant rough sets is put forward here,taking the Shuangtaizi estuarine wetland as the research object divided into 8 categories （water,culture pond,suaeda,reed,paddy,beach,residential land and mixed vegetation） and Landsat-8 OLI remote sensing data on 26 May 2014 as the data source. To satisfy the application demand,image preprocessing was needed including radiometric correction at the systematic level and geometric correction by ground control points and digital elevation model data for Landsat-8 data products. First,deal with sample data set collected in the study area using tolerant rough set theory preprocessing to obtain new training samples. Second,a classification model based on BP neural network was established by Matlab software platform,and land-cover information was extracted. Then it conducted evaluation for classification effect using a confusion matrix. Results show that the BP neural network based on tolerant rough set classification can eliminate noise in the training sample data,improve the success rate of training of the network and shortening the network training time. It obtained a good classification effect with an overall accuracy of 91.25%,Kappa coefficient 0.8969,increased 7.92% and 0.0926 than traditional BP neural network classification method,and 3.03% and 0.0357 higher than the pretreatment method based on rough set attribute reduction. BP neural network classification based on tolerant rough set is a preferable land-cover classification method which can help managers monitor wetland dynamics.