A Hyperspectral Remote Sensing Image Classification Method Using Atrous Convolution

Abstract

The invention relates to the field of deep learning and image classification processing of artificial intelligence, in particular to the classification of ground features of hyperspectral images by using a deep convolutional neural network. In the model training stage, the hyperspectral image to be classified is first converted into a single-pixel multi-dimensional image with a label, and the feature map of a single pixel is extracted by using a combination model based on atrous convolution, and then a three-convolution layer, three The convolutional neural network of the activation layer extracts the main features, and finally uses a layer of full connection and Softmax function as a classifier to complete single-pixel classification. The invention makes full use of the spectral information of the hyperspectral remote sensing image, solves the gridding problem in the hollow convolution, expands the convolution receptive field, improves the convolution efficiency, and improves the classification accuracy while reducing the amount of parameters. An Efficient and Lightweight Deep Atrous Convolutional Neural Network for Hyperspectral Remote Sensing Image Classification with Pixel Accuracy.

Description

[0001] Aiming at the problems of dimensional "disaster" and insufficient utilization of spectral feature information in hyperspectral remote sensing image classification using deep convolutional network (DCNN), the atrous convolution structure of deep convolutional neural network is applied to hyperspectral remote sensing image lossless spectral domain features In the study, invented a method of pixel-level classification of hyperspectral remote sensing images using the lightweight network model of hole convolution, which solved the gridding problem in hole convolution, expanded the receptive field of deep convolution network, and improved It improves the convolution efficiency, reduces the number of parameters, and maintains a high classification accuracy. technical field [0002] The invention relates to the field of deep learning and image classification processing of artificial intelligence, in particular to the classification of ground features of hyperspectral images by u...

AI Technical Summary

Problems solved by technology

[0008] 1) The image target classification model mainly learns spatial features at different scales through convolution and pooling cascades. Hyperspectral image datasets often have small spatial dimensions, so the deep learning model is used directly to classify hyperspectral images. Its spatial features are insufficient, and there will be Hughes phenomenon in the learning process, that is, the contradiction between the small number of samples and the hyperspectral dimension

[0009] 2) Existing deep learning methods generally use dimensionality reduction for hyperspectral images first, select some dimensions that can represent the main features of objects, and then send them to the classic deep learning model for training, and then obtain classification results, but the dimensionality reduction process It will cause a certain loss of spectral information, which is not conducive to fine classification

Images