基于小样本学习的高光谱遥感图像分类算法

HSI Classification Algorithm Based on Few-Shot Learning

高光谱图像(Hyperspectral Image,HSI)具有丰富的地物空间信息和光谱信息,为地物的精细分类提供了有利条件.当前高光谱遥感图像分类方法,在标记样本较少的情况下,其分类性能相对有限.那么,如何充分发挥小样本学习领域的相关理论与方法,以提升高光谱遥感图像分类性能,是一个意义重要且具有挑战性的问题.为此,本文以标记的小样本为背景,对现有基于深度学习的高光谱遥感图像分类算法进行改进,其核心思路在于:利用基于深度三维神经网络的嵌入模型从训练数据集中学习投影函数,进而将视觉特征映射到语义表示,同时通过最小化损失函数使其类内间距较小、类间间距较大,再使用训练好的网络嵌入模型将每个未标记的目标类表示在相同的嵌入空间中,最后,分别选用欧氏距离和卷积神经网络作为距离度量方式,计算嵌入空间中嵌入与类原型之间的距离,从而实现投影后的样本分类.本文在印第安松、萨利纳斯和帕维亚大学数据集上进行了实验,其结果表明改进的算法具有较好的高光谱遥感图像分类性能.

Hyperspectral Image(HSI)contain rich spatial and spectral information,which provides favorable conditions to identify different classes of terrain accurately.At present,the performance of hyperspectral image classification method is relatively undesired with few labeled samples.Therefore,it is an important and challenging problem to give full play to the relevant theories and methods in few-shot learning to improve the classification performance of hyperspectral image classifiers.For this reason,based on few samples labeled,this paper improved the existing hyperspectral image classification algorithm based on deep learning.Above all,an embedded model based on the deep 3D-CNN is used to learn the projection function from the training datasets,which maps the visual features to the semantic representation,and make samples from the same class close and those from different classes far by minimizing the loss function.Then the pretrained network is employed to represent each unlabeled samples in the same embedding space.Finally,the Euclidean distance and CNN are respectively used as distance metric to obtain the distance between the embedding feature and the class prototype,so as to classify the samples in the test datasets.Experiments are carried out on the Indian pines,Salinas and Pavia University,and the experimental results show that both the improved algorithm based on deep learning and algorithm based on relation network can achieve the desired classification performance.