A hyperspectral remote sensing data classification method based on depth neural network

Abstract

A hyperspectral remote sensing data classification method based on depth neural network comprises the following steps: (1) reading hyperspectral data; (2) determining the number of categories and selecting training samples and test samples; (3) performing Spatial and spectral feature joint extraction based on three-dimensional convolution and space pyramid pooling; (4) Establishing the hyperspectral data classification model based on depth neural network; (5) optimizing the model by transfer learning strategy and avoiding over-fitting; (6) inputting All the sample points to be classified intothe model for classification, and the classification result map is obtained. The classifier model in this method migrates the features from the source domain data to the target domain data by optimizing the training of migration learning strategy. Therefore, this method can obtain higher classification accuracy when the number of training samples is small.

Description

technical field [0001] The invention relates to a hyperspectral remote sensing data classification method based on a deep neural network, which belongs to the field of hyperspectral data processing methods and application technologies, and is suitable for the theoretical method and application technology research of hyperspectral data supervised classification. Background technique [0002] In the classification of hyperspectral remote sensing data, the classification accuracy of hyperspectral remote sensing images will show the Hughes phenomenon that first increases and then decreases as the spectral dimension increases. For the spectral dimension of hyperspectral images, it is often impossible to provide enough training samples for classifiers, so the small sample problem greatly affects the accuracy of hyperspectral remote sensing image classification. At present, there are usually three solutions to the small sample problem of hyperspectral image classification: the firs...

AI Technical Summary

Problems solved by technology

In hyperspectral image classification, the Hughes phenomenon is one of the important problems that affect the results of hyperspectral image classification. The key is that the number of labeled samples is small. However, obtaining sufficient labeled samples in practical applications requires a high price.

The extraction of spectral information and spatial information is a key step in hyperspectral data classification based on deep neural networks, but most of the existing deep neural network classification models use two-channel feature extraction, that is, the spectral channel extracts spectral information and the spatial channel Extract spatial information, this feature extraction method cannot realize the extraction of spatial information and spectral information in a single channel

Images