联合一二阶池化网络学习的遥感场景分类

目前大多数池化方法主要是从一阶池化层或二阶池化层提取聚合特征信息，忽略了多种池化策略对场景的综合表示能力，进而影响到场景识别性能。针对以上问题，提出了联合一二阶池化网络学习的遥感场景分类模型。首先，利用残差网络ResNet-50的卷积层提取输入图像的初始特征。接着，提出基于特征向量相似度的二阶池化方法，即通过特征向量间的相似度求出其权重系数来调制特征值的信息分布，并计算有效的二阶特征信息。同时，引入一种有效的协方差矩阵平方根逼近求解方法，以获得高阶语义信息的二阶特征表示。最后，基于交叉熵和类距离加权的组合损失函数训练整个网络，从而得到富于判别性的分类模型。所提方法在AID（50%训练比例）、NWPU-RESISC45 （20%训练比例）、CIFAR-10和CIFAR-100数据集上的分类准确率分别达到96.32%、93.38%、96.51%和83.30%，与iSQRT-COV方法相比，分别提高了1.09个百分点、0.55个百分点、1.05个百分点和1.57个百分点。实验结果表明，所提方法有效提高了遥感场景分类性能。

Joint 1-2-order pooling network learning for remote sensing scene classification

At present， most pooling methods mainly extract aggregated feature information from the 1-order pooling layer or the 2-order pooling layer， ignoring the comprehensive representation capability of multiple pooling strategies for scenes， which affects the scene recognition performance. To address the above problems， a joint model with first- and second-order pooling networks learning for remote sensing scene classification was proposed. Firstly， the convolutional layers of residual network ResNet-50 were utilized to extract the initial features of the input images. Then， a second-order pooling approach based on the similarity of feature vectors was proposed， where the information distribution of feature values was modulated by deriving their weight coefficients from the similarity between feature vectors， and the efficient second-order feature information was calculated. Meanwhile， an approximate solving method for calculating square root of covariance matrix was introduced to obtain the second-order feature representation with higher semantic information. Finally， the entire network was trained with the combination loss function composed of cross-entropy and class-distance weighting. As a result， a discriminative classification model was achieved. The proposed method was tested on AID （50% training proportion）， NWPU-RESISC45 （20% training proportion）， CIFAR-10 and CIFAR-100 datasets and achieved classification accuracies of 96.32%， 93.38%， 96.51% and 83.30% respectively， which were increased by 1.09 percentage points， 0.55 percentage points， 1.05 percentage points and 1.57 percentage points respectively， compared with iterative matrix SQuare RooT normalization of COVariance pooling （iSQRT-COV）. Experimental results show that the proposed method effectively improves the performance of remote sensing scene classification.