Hierarchical nonlinear subspace dictionary learning method for remote sensing image scene classification

Abstract

The invention discloses a hierarchical nonlinear subspace dictionary learning method for remote sensing image scene classification. HNSDL projects a data sample into a series of subspaces by using a hierarchical nonlinear method and realizes data dimension reduction; in order to improve the identification capability of the model, a local structure constraint term of sparse coding is introduced in the learning process. The method has the following technical progress that: (1) different from the existing linear method of a linear transformation matrix, the proposed method maps data to a series of projection spaces through a series of nonlinear functions; compared with a linear transformation matrix, the nonlinear projection can better represent the nonlinear structure of the remote sensing image. And (2) by utilizing a local structure constraint item with label information, dictionary learning not only retains a nonlinear local structure of data, but also enables intra-class samples to be compact, and the class separability of the dictionary is enhanced. And (3) projection space learning and dictionary learning are simultaneously learned in one model, and an alternative iterative optimization method is used for parameter optimization, so that the parameters of the model can simultaneously obtain optimal solutions.

Description

technical field [0001] The invention relates to the field of remote sensing image processing, in particular to a hierarchical non-linear subspace dictionary learning method for scene classification of remote sensing images. Background technique [0002] With the rapid development of many fields such as computer science, remote sensing, and communication engineering, remote sensing images have exploded, making large-scale surface monitoring possible. Automatic land cover classification based on remote sensing images has important application value in urban planning. Land cover classification of remote sensing images is usually based on supervised learning methods, and many machine learning models, such as support vector machines, Markov random fields, random forests, etc., have been widely used in hyperspectral image classification. However, these methods only consider the spectral features of remote sensing images, while ignoring the sparsity and spatial information hidden ...

AI Technical Summary

Problems solved by technology

Second, training a dynamic learning model often requires a large number of annotated samples, however, it is difficult and time-consuming to annotate a large number of remote sensing images in new scenes

Third, the high spatial resolution and rich spectral information of remote sensing images make it difficult to directly construct deep learning models for classification

However, the shortcomings of the methods proposed so far are that they all perform dictionary learning in the original space of samples. First, due to the high-dimensional characteristics of remote sensing images, the memory usage and computational complexity in dictionary learning increase sharply with the increase in dimensionality. Performance also suffers from the curse of dimensionality

Second, for complex problems such as classification of remote sensing images, the nonlinear relationship between samples cannot be well utilized in the original space of samples

Although they are easy to implement, they may not fully preserve the best properties of dictionary learning, since most of these methods just perform dimensionality reduction and dictionary learning as two separate stages

Images