PTL-CFS based deep convolutional neural network model for remote sensing classification

Processing high-dimensional remote sensing images data with conventional convolutional neural networks raises certain issues such as prolonged model convergence time, vanishing gradient, convergence of the non-minimum values, etc. due to its high time-complexity and random initialization parameters nature. Aiming at those issues, this article proposes a convolutional neural network remote sensing classification model based on PTL-CFS. This model approach first utilizes parameter transfer learning algorithm to obtain the CNN initialization parameters of the target area, then it uses correlation-based feature selection algorithm to eliminate the redundant features and noises from the original feature set, finally, it classifies the remote sensing images using a conventional CNN model. This article has proven the validity of such network model when classifying remote sensing images in the Zha long wetland, Heilongjiang. Experiments show that the addition of PTL can accelerate the loss function of the convergence rate in CNN. The algorithm combined with CFS algorithm, compared with other algorithms to reduce the algorithm execution time and get better classification accuracy.