Research On Classification Of Hyperspectral Remote Sensing Images Of Coastal Wetlands Based On Deep Learning CNN

Coastal wetland is an important ecosystem,and remote sensing technology is an important method for monitoring coastal wetlands.Hyperspectral remote sensing data has the characteristics of the combination of the spectrum and the image,which provides the possibility for the fine classification of remote sensing images.Convolutional neural network is an important means of image classification,and it is a new direction of remote sensing image fine classification.In view of the slow convergence rate of gradient descent algorithm in traditional CNN model,this paper develops a CNN model based on conjugate gradient method.The conjugate gradient algorithm is used to update the descending gradient of the model,which overcomes the shortcomings of slow convergence and low precision of the CNN model when there are many training samples.To solve the problem of overfitting or underfitting caused by the fixed learning rate in the traditional CNN model,this paper proposes a self-adaptive learning rate CNN model combining golden section method and quadratic interpolation method.The model can automatically adjust the learning rate,instead of manually setting the learning rate manually,which can significantly reduce the workload of the traditional model to find the optimal adjustment parameters.In this paper,a conjugate gradient method-based CNN model is developed and a CNN model of adaptive learning rate is proposed.The classification experiment is carried out using the 2-view coastal wetland hyperspectral imagery.The main conclusions are as follows:1)The conjugate gradient method-based CNN model converges faster when the computational load of the model increases.When a certain accuracy requirement is reached,the time consumption is significantly less than the traditional CNN model,only 1/3 of the traditional CNN model.The conjugate gradient method based CNN model shows the anti-jamming ability to noise.Under the condition of increasing the number of training samples and different levels of noise interference,compared with the traditional CNN model,it can provide higher classification accuracy.2)The conjugate gradient method has the advantage of stable convergence in solving large-scale nonlinear optimization problems.The conjugate gradient method-based CNN model can still ensure the classification accuracy when the number of training samples is large,and it is suitable for high data volume hyperspectral.Image classification.Increasing the number of batch training samples,the accuracy of the decline in the conjugate gradient descent CNN model is significantly smaller than that of the CNN model.And the accuracy of batch training samples increased to 160 can still be maintained at more than 80%,higher than the CNN model 2.8 %.3)The CNN model of adaptive learning rate is affected little by the initial value.When the learning rate search interval is fixed,the classification accuracy of the traditional CNN model is greatly influenced by the initial value.The classification accuracy of the floating value of up to 10.8%,while the accuracy of the proposed model's classification accuracy is basically stable within 1.2%,significantly lower than the traditional CNN model,which achieves adaptive adjustment of the learning rate.When the range of learning rate is controlled between 0 and 1,the proposed model can achieve better classification accuracy.