Classification method based on deep migration learning and neighborhood noise reduction

The invention discloses a classification method based on deep migration learning and neighborhood noise reduction. The method comprises steps of migrating CNN shallow network weight parameters pre-trained on the source data set to a target data set; through fine tuning over a network, initializing CNN deep network weight parameters trained by a target data set network randomly; training again onthe target data set to finish hyperspectral image classification based on transfer learning, then performing optimal neighborhood point noise reduction based on eight neighborhood point mode tags onan image marking result of hyperspectral image classification output through transfer learning, and finally outputting a noise-reduced image classification result. 本发明公开了一种基于深度迁移学习与邻域降噪的分类方法，将在源数据集上预训练的CNN浅层网络权值参数迁移至目标数据集，通过网络微调，随机初始化目标数据集网络训练的CNN深层网络权值参数，并在目标数据集上重新训练，完成基于迁移学习的高光谱图像分类，然后，再对通过迁移学习输出的高光谱图像分类的图像标记结果进行基于八邻域点众数标签的最优邻域点降噪，最终输出降噪后的图像分类结果。