Research On Optimization Of Machine Learning Model For Urban Functional Area Classification

With the advancement of China's urbanization and the construction of smart cities,new challenges have been raised for refined urban governance.A deep understanding of urban regional functions and a grasp of urban spatial structure are of great significance to improving the level of urban governance.This article focuses on the classification of urban functional areas through intelligent image recognition technology.Based on a dataset of 15,000 remote sensing images of urban functional areas and corresponding text information,this paper analyzes multiple theoretical methods,and compares the pros and cons of experiments to verify and improve the algorithm model.Finally,in the classification task for different size data sets,the PCA-SVM algorithm model optimized by PSO and the improved ResNet50 network model were used for classification,The main content of this article is:1.Using a variety of data enhancement algorithms such as spatial transformation,adding Gaussian noise,color enhancement,sharpening,etc.,to enhance the data of the five types of datasets of a total of 15,000 remote sensing images of urban functional areas,to avoid overfitting to a certain extent Occurrence,improves the stability and robustness of the model,and improves the generalization ability of the model.2.Based on the data preprocessing,proposed a PCA-SVM-based urban functional area partitioning model.The PCA algorithm is used to extract image features,which are sent to the SVM classifier using a one-to-many method for classification,and the PSO algorithm is used to solve the parameter finding.To optimize the network model.The accuracy of the optimized network model is increased by 1.20%,and the kappa coefficient is increased by 0.0206.However,the accuracy of the model will decrease with the increase of the data set.3.Based on 15,000 remote sensing data sets of urban functional areas,through theoretical analysis,the VGG-16 network model,ResNet34 network model,and ResNet50 network model were selected for experimental comparison.Based on experimental verification results,the ResNet50 network model has the highest accuracy rate and relatively little increase in training time.Finally,the ResNet50 network model was initially selected as the urban functional area partition model.4.Based on the preliminary selection of the urban functional area partitioning model,the ResNet50 network model is improved.First,the mixup algorithm is introduced on the basis of the original data samples to generate virtual samples,which enhances the robustness of the network and improves the stability of model training.Secondly,text features are added and feature stitching is performed through a fully connected layer.The improved network calculation does not increase significantly,and the test accuracy rate based on the urban functional area dataset is increased by 1.31%to 72.30%.5.Optimization experiments were performed based on the improved ResNet50 network model.Through multiple experiments and comparative analysis,the batch size,activation function,and gradient descent algorithm of the urban functional area partitioning model were determined,which further improved the classification accuracy.The final accuracy reached 72.66%and the kappa coefficient reached 0.7249.6.Create a user interface based on the final network model which makes the classification of urban functional areas more convenient.In this paper,a variety of algorithms are used for experimental comparison and analysis.The PSO-optimized PCA-SVM algorithm is used on the data set of the smaller urban functional area,with an accuracy rate of 70.80%;in the case of a large data set,the ResNet50 network model is used Improvement,and the final accuracy rate reached 72.66%,which can better complete the task of dividing urban functional areas.