特征重标定网络的高光谱影像分类方法

高光谱影像特征的利用率对提高其分类精度具有重要意义。为充分利用影像的特征,提出了一种特征重标定网络的高光谱影像分类方法。该方法通过全局平均池化将特征图转换为具有全局信息的实数,利用全连接层与非线性层生成能够代表各通道相对重要性的权值,进而采取加权法完成初始特征的重标定。为验证该方法的有效性,选取PaviaU和KSC两组高光谱影像数据进行实验。结果表明,提出方法总体分类精度分别达到98.38%和95.61%,可为高光谱影像提供有效的类别判定特征,有助于提高影像分类精度并获取平滑的分类结果图。

Feature Recalibration Network for Hyperspectral Imagery Classification

It is essential to make full use of the abundant features in hyperspectral image for the classification. The method for hyperspectral image classification is proposed based on feature recalibration network. With global average pooling, the feature maps can be transformed to values representing global receptive field. Then the full connected layers and the nonlinear layers are used to generate the weights representing the importance of each feature channel. At last, the weighted method is applied to recalibrate the original feature. The PaviaU and the KSC hyperspectral data sets are used to verify the proposed method. The overall classification accuracies separately reached 98.38% and 95.61% based on the feature recalibration network. The results show that the proposed approach can provide the effectively distinguished features, and conduce to improve the accuracy of imagery classification and obtain the smoother classification maps.