Hyperspectral space-spectrum joint feature extraction method based on transfer learning

Abstract

The invention discloses a hyperspectral space-spectrum joint feature extraction method based on migration learning, which belongs to the field of deep learning remote sensing. The method of extracting the space-spectrum joint features of hyperspectral data is to first design 1D CNN and 2D CNN to extract the spectral and spatial features of hyperspectral data, and then fuse the two parts of features. In order to overcome the contradiction that the deep neural network needs a large amount of training data and the hyperspectral data lacks labeled samples, the present invention adopts the method of migrating the model ResNet-18 pre-trained on the RGB image data set ImageNet to the hyperspectral image target domain to realize network parameter sharing , to reduce the computational cost of training the model. Based on the extracted combined features, the SoftMax layer is trained to implement the hyperspectral target classification task. Finally, by fine-tuning the transfer learning strategy, the transferred model is more suitable for hyperspectral data and the classification accuracy is improved. The invention has a clear structure, is easy to implement, and has profound theoretical foundation and practical significance.

Description

technical field [0001] The invention mainly relates to the field of hyperspectral image classification, in particular to a hyperspectral data space-spectrum joint feature extraction method based on migration learning, which is used for hyperspectral classification. Background technique [0002] With the development of spectral imaging technology, hyperspectral imaging has attracted extensive attention in the field of remote sensing due to its broadband detection from visible light to near-infrared wavelength range. Among them, the application of target detection and image classification for hyperspectral data is becoming more and more mature. Early hyperspectral classification methods largely relied on human expertise, and the classification basis was mostly based on shallow features, resulting in low classification accuracy and high cost. With the development of deep learning in recent years, deep learning methods have better feature extraction and expression capabilities ...

AI Technical Summary

Problems solved by technology

[0009] The main technical problem to be solved by the present invention is: for the traditional hyperspectral image classification data labeling cost is high, the source of the data set is few, and the training time is long, etc., the present invention provides a clear structure, easy to implement, good classification effect and short time-consuming The hyperspectral classification method of

Images