A hyperspectral image classification method based on adaptive spatial-spectral multi-scale network

Abstract

The invention discloses a hyperspectral image classification method based on an adaptive space-spectrum multi-scale network, which includes a training stage and a prediction stage. The training phase includes image preprocessing, sample selection and network training. First, the hyperspectral image is subjected to dimensionality reduction processing. In the sample selection stage, an appropriate proportion of labeled samples is randomly selected for each class from the original and dimensionality-reduced hyperspectral images, and then the designed network is used for training. In the prediction stage, the entire image is directly input into the network to obtain the final classification result.

Description

technical field [0001] The invention belongs to the technical field of remote sensing image processing, in particular to an adaptive space-spectrum multi-scale network method. Background technique [0002] With the development of sensor technology, it is possible to obtain hyperspectral images with hundreds of channels per pixel, which contain very rich information. Land cover classification of hyperspectral images has been a hot topic in recent years. Its goal is to assign a unique semantic label to each pixel in hyperspectral images, and then generate an accurate and complete classification map. This classification map can guide decision makers in industries such as agriculture, environmental monitoring, materials analysis, and more. However, due to the complexity of the spectral and spatial structure of the hyperspectral image itself, this task is still challenging. [0003] Traditional classification methods directly input raw spectral vectors into the classifier. Thi...

AI Technical Summary

Problems solved by technology

[0005] However, the features finally extracted from the above work are often of a single scale, which means that the size of the receptive field of each pixel in the feature map input to the classifier is the same, which limits the improvement of classification accuracy, and no matter how the ground object is scaled, its The category should be invariant, so features of different scales should be taken into account, which helps to improve the robustness of the model

In addition, after the development of deep learning, there have been a lot of work using the space-spectrum fusion network for hyperspectral classification. However, there is a commonality in these works in the space part, which is to rely solely on the ability of CNN's local perception. In the convolution operation of CNN, Adjacent pixel values ​​are processed in parallel by point multiplication, which means that it is difficult for CNN to perceive the spatial context information existing between adjacent pixels, which limits the improvement of classification accuracy

Images