Determination Method of Key Convolutional Layer Hyperparameters of Convolutional Neural Networks for Remote Sensing Classification

Abstract

The invention provides a method for determining key convolutional layer hyperparameters of a remote sensing classification convolutional neural network, comprising the following steps: construction of a convolutional neural network sample set; construction of a convolutional neural network structure; key layer hyperparameters of a convolutional neural network Determination: Select one of the convolutional layers as the key layer, preset the size of the convolution kernel of the key layer, and calculate the convolution scale; calculate the convolution step size according to the set rules according to the convolution kernel and convolution scale of the key layer; Set the convolution kernel of other convolutional layers as kernelsize, and the convolution step of other convolutional layers as 1; the subsequent downsampling size adopts mean downsampling or maximum downsampling. The invention proposes the concept of convolution scale based on the image input size and convolution kernel size, which is compatible with the remote sensing space scale, and on this basis, provides a joint determination key based on the input size and convolution scale The method of layer hyperparameters can reduce the time required for algorithm tuning and improve the classification accuracy of object-oriented remote sensing.

Description

technical field [0001] The invention relates to the field of remote sensing classification, in particular to a method for determining key convolutional layer hyperparameters of a remote sensing classification convolutional neural network. Background technique [0002] Geographic target recognition through remote sensing images is an important link in the application of remote sensing technology to practical problems. Whether it is thematic information extraction, dynamic change monitoring, thematic mapping, or remote sensing database construction, remote sensing image classification technology is inseparable. Object-oriented remote sensing technology gathers many adjacent pixels to form a spatial image object with certain semantics. It integrates the spatial, texture and spectral information contained in remote sensing data. Its core is image object construction and image object classification. Convolutional neural network is a deep learning algorithm. One of them, the algor...

AI Technical Summary

Problems solved by technology

However, when the convolutional neural network is applied to remote sensing image classification, there is still a problem of how to determine the hyperparameters (convolution kernel size and step size) of the convolutional neural network according to the input remote sensing images of different resolutions.

Images