【摘要】 针对喀斯特地区受山区复杂地形的影响岩溶山峰在太阳辐射下存在阴坡和阳坡,高分辨率遥感图像中喀斯特森林植被的光谱特征较复杂,"同物异谱"和"异物同谱"现象严重,导致传统的机器学习算法提取森林植被精度不高的问题,根据实践经验将多源多特征融合构成提取喀斯特森林信息特征,改进标准的卷积神经网络(convolutional neural network,CNN),将支持向量机(support vector machine,SVM)与卷积神经网络相结合(CNN-SVM)应用于遥感分类,并与CNN、随机森林(random forest,RF)、支持向量机等方法进行比较。结果表明,CNN-SVM、CNN两种深度学习方法的提取喀斯特森林信息精度均明显高于RF和SVM等浅层模型方法。CNN-SVM综合了CNN提取遥感高阶特征的能力和SVM的分类性能,分类精度在90%以上,高于标准的CNN。深度学习CNN可有效地区分农作物,提高喀斯特森林植被信息的提取精度。更多还原

【Abstract】 As there are shaded slopes and sunny slopes resulted from the hilly effects of complex terrain under solar radiation in Karst regions, the spectral characteristics of Karst forest are more complex in high resolution remote sensing images. It is very difficult to extract forest accurately from remote sensing images using traditional machine learning algorithms because of serious phenomena of different spectra with same object and same subject with different spectra. Firstly, the Karst forest extraction features were got from multi-feature fusion from multi-source remote sensing data. Then a hybrid convolutional neural network and support vector machine(CNN-SVM) classifier designed by replacing the last output layer of CNN with an SVM classifier was proposed for land use classification with high resolution remote sensing images. The classified accuracy with CNN-SVM was compared to those with CNN, SVM and RF(random forest). Results show that deep learning classifiers, such as CNN-SVM and CNN, show their superiority over RF and SVM traditional non-deep classifiers in extracting Karst forest. As CNN worked as a trainable high-level feature extractor and SVM performed as a recognizer in CNN-SVM, the classification accuracies of high resolution remote sensing images with CNN-SVM classifier are over 90%. CNN-SVM classifier has better performance than CNN implementations. Deep learning CNN classifier can be used to distinguish forest from crops with higher accuracy.更多还原