基于多尺度特征融合的遥感图像场景分类

摘要

为了解决遥感图像场景分类中因样本量小而分类精度不高的问题, 提出了一种基于多尺度特征融合 (MSFF) 的分类方法.首先, 对遥感图像进行尺度变换, 得到同一遥感源图像的多个不同尺度图像.接着, 将其分别输入深度卷积神经网络 (DCNN) 中进行卷积操作.然后, 将各卷积层和全连接层提取出的不同尺度特征进行降维和编码/平均池化操作.最后, 将各尺度特征进行编码融合并利用多核支持向量机 (MKSVM) 进行场景分类.在两个公开遥感图像数据集UCM Land-Use和NWPU-RESISC45中进行试验, 分类精度最高分别达到98.91％和99.33％.本文方法能够利用不同尺度的图像特征, 结合低、中、高层语义表示, 使融合特征的可辨识性更高, 同时使用多核支持向量机提高了深度网络学习的泛化能力, 因此分类效果更好.%To solve the low accuracy problem of remote sensing image scene classification due to small sample sizes, a classification method was proposed based on Multiscale Features Fusion (MSFF) .First, the remote sensing images were scaled to obtain several different scale images of the same remote sensing image.Thereafter, they were inputted into a Deep Convolutional Neural Network (DCNN) for convolutional operation, and the different scale features of the convolutional and the fully connected layers were reduced and coded or average pooled.Finally, the different scale features were coded and fused, and a multikernel support vector machine was used to classify the scenes.In the two public remote sensing image data sets UCM Land-Use and NWPU-RESISC45, the highest classification accuracy of the experiment are 98.91％ and 99.33％, respectively.This method can use image features of different scales and low, middle and high-level semantic representations combined, thus the fusion feature is more recognizable.Furthermore, the use of a multikernel support vector machineimproves the generalization of the deep network learning ability, so the classification effect is better.