Improved high-resolution remote sensing image classification method based on deep learning

Abstract

The invention discloses an improved high-resolution remote sensing image classification method based on deep learning. On the basis of the deep learning theory, a seven-layer convolutional neural network is designed; a high-resolution remote sensing image sample is inputted into the network to carry out network training and last two full connection layers obtained by learning are outputted as twodifferent high-level features of the remote sensing image; dimension reduction is carried out by using a principal component analysis for the output of the fifth pooling layer of the network, whereinthe result after dimension reduction is used as a third high-level feature of the remote sensing image; the three kinds of high-level features are fused in series; and then an effective logistic-regression-based classifier is designed to classify the remote sensing image. According to the invention, feature extraction is carried out on the high-resolution remote sensing image based on the deep learning theory and the features obtained by learning have high expressive force and robustness. Besides, the extracted high-level features are fused and the fused feature is inputted into the logistic regression classifier, so that the good classification result is obtained.

Description

technical field [0001] The invention belongs to the field of image processing, and in particular relates to an improved classification method of high-resolution remote sensing images based on deep learning. Background technique [0002] In the field of optical remote sensing image processing, the spatial resolution of remote sensing images refers to the minimum size of a single feature or two adjacent features that can be identified by optical sensors. The higher the spatial resolution, the richer the morphological information of ground features contained in remote sensing images. The spatial resolution of optical remote sensing satellites currently in commercial operation has reached the "sub-meter level". For example, the US WorldView-4 satellite launched in 2016 can provide high-definition ground images with a resolution of 0.3m. In recent years, with the rapid development of my country's space technology, especially the implementation of major projects of high-spatial-r...

AI Technical Summary

Problems solved by technology

[0009] (1) The existing remote sensing image classification methods often require remote sensing image samples to be preprocessed, and the preprocessing process is easily affected by many external factors, such as atmospheric conditions, surface environment and other factors, which will weaken the classification performance

[0010] (2) Existing remote sensing image classification methods require too much human intervention when extracting features. Different image feature extraction algorithms are selected to extract various types of image features for subsequent image classification. The feature extraction process is complicated.

[0011] (3) The features extracted by existing remote sensing image classification methods are not robust enough

Existing classification methods often extract the underlying features of images, such as texture features, color features, and contextual prior information. These features may also have redundancy, and perform poorly in terms of displacement and rotation, resulting in low classification accuracy.

Images