The deep belief network model, which is widely used in deep learning, consists of a multi-layer constrained Boltzmann machine and a back-propagation network. The authors have conducted parameter sensitivity experiments on the number of iterations, the number of hidden layers and the number of hidden layer nodes in the DBN network for remote sensing image classification, and obtained a set of optimal parameter setting schemes. Moreover, the DBN algorithm has been enhanced with an improved Dropout strategy. The improved Dropout strategy selects only part of the data to clear the weight at a time, and a local area randomly clear strategy is adopted, which will save the local information of the image itself, and enhance the generalization ability of the model. In order to verify the advantages of the improved DBN algorithm model, the classification results of DBN, KNN, random forest and SVM have been compared. And the results show that classification accuracy of the improved DBN has been greatly improved, which is increased by about 2.5% compared to DBN. The improved DBN classification results are processed then, including connected areas marking, noise removal, morphological transformation and edge extraction, and the boundary information of the building is obtained according to the target shape characteristics. Finally, the experiment on the morphological characteristics of the building also shows it can extract better edge information of the building.

在深度学习中广泛使用的深度信念网络模型由多层约束的Boltzmann机器和反向传播网络组成。作者对遥感图像分类的DBN网络中的迭代次数，隐藏层数和隐藏层节点数进行了参数敏感性实验，并获得了一组最优的参数设置方案。此外，DBN算法已通过改进的Dropout策略进行了增强。改进的Dropout策略一次只选择部分数据来清除权重，采用局部随机清除策略，可以节省图像本身的局部信息，增强了模型的泛化能力。为了验证改进的DBN算法模型的优势，比较了DBN，KNN，随机森林和SVM的分类结果。结果表明，改进后的DBN的分类精度得到了极大的提高，与DBN相比提高了约2.5％。然后处理改进的DBN分类结果，包括连接区域标记，噪声消除，形态转换和边缘提取，并根据目标形状特征获得建筑物的边界信息。最后，对建筑物形态特征的实验还表明，它可以提取建筑物的较好边缘信息。然后处理改进的DBN分类结果，包括连接区域标记，噪声消除，形态转换和边缘提取，并根据目标形状特征获得建筑物的边界信息。最后，对建筑物形态特征的实验还表明，它可以提取建筑物的较好边缘信息。然后处理改进的DBN分类结果，包括连接区域标记，噪声消除，形态变换和边缘提取，并根据目标形状特征获得建筑物的边界信息。最后，对建筑物形态特征的实验还表明，它可以提取建筑物的较好边缘信息。