Research on vgg16 convolutional neural network feature classification algorithm based on Transfer Learning

feature classification has broad development prospects in remote sensing applications. In order to accurately identify feature categories, this paper constructs a feature classification model with vgg16 neural network as the core network, and realizes the classification of water, farmland, buildings, roads and trees for light and small SAR images through in-depth learning route. Firstly, the algorithm constructs a new data set, and each picture in the data set is a feature category with labels. Secondly, the vgg-16 framework of deep convolution neural network on the new data set is built. According to the small scale of the data set, the pre training model obtained by migration learning is introduced, and the convolution layer of vgg16 network is locally adjusted according to the body characteristics and natural scenes of various ground objects, so as to optimize the main model parameters, so as to realize the prediction of water, farmland, buildings, roads Accurate classification of trees. The experimental results show that the average accuracy of vgg16 network without parameter adjustment can reach 75%, the average accuracy of vgg16 network after optimizing model parameters can reach 81%, and the average accuracy of feature classification after adding pre training model is 87.5%.