Remote Sensing Image Classification Based On Transfer Learning

Accurate and efficient classification of remote sensing images has always been a hot research direction in the field of remote sensing applications.Its applications include forest fire detection,surface cover identification,geographic image retrieval and urban and rural planning.The rapid development of remote sensing technology has produced a large number of remote sensing image sets that need to be classified.This puts higher requirements on the classification methods of remote sensing images.The traditional classification method can’t achieve the ideal classification accuracy by extracting low-and medium-level image features.Accuracy rate remote sensing image classification method is an important application value.Convolutional neural networks in the field of deep learning currently perform well in image classification and target detection tasks,and their classification accuracy is much higher than that based on traditional manual design features.But convolutional neural networks require a large number of label data sets to be trained.For models with good classification effect,remote sensing images are difficult to obtain large-scale label datasets due to the complicated process of cropping and labeling.The use of remote sensing image sets with small sample size can train the neural network model to be over-fitting,and use remote sensing.The network model built and trained by the data set has poor reusability,and it has a poor classification effect on the new remote sensing image set except the training data set.In order to solve the above problems,this paper proposes a migration algorithm framework based on migration learning method and pre-training convolutional neural network model,and applies the pre-trained model and its weight parameter migration on large-scale dataset to small sample remote sensing image classification.The main research contents of this paper are as follows:1)This paper first introduces the main problems of remote sensing image classification,introduces the concept of transfer learning.The concept transfer method details the three pre-trained convolutional neural network model structures used in the experiment.2)Constructing a transfer framework model suitable for small sample remote sensing datasets,and migrating the neural network pre-trained on large-scale datasets and its parameters to small sample remote sensing data by combining transfer learning and convolutional neural networks.Set up and then fine-tune the training.Experiments are carrying out on the basis of this framework,and multiple sets of contrast experiments are conducted to study the effects of different dataset sizes,different network structures and different fine-tuning depths on the classification of transfer learning,and the maximum transfer learning for remote sensing image datasets is summarized.3)A solution algorithm for fine-tuning the optimal depth of the transfer model based on peak convergence is proposed.The algorithm greatly reduces the network training time and finds the transfer time for the transfer while finding the optimal fine-tuning depth and the optimal classification accuracy.The model seeks to fine tune the optimal depth and proposes a benchmark strategy.4)Finally,a feature fusion algorithm based on transfer learning pre-trained model is proposed.The features of the three pre-trained models extracted from the input image and the HOG features of the image are merged.The fusion feature is used to train the linear classifier to obtain the classification.As a result,multiple sets of contrast experiments were set up.The experimental results show that the classification accuracy of the feature fusion algorithm is significantly higher than that of the single model participating in the fusion,which proves the superiority of the algorithm.