The Optimizing Of Object-based Random Forest Classification In Remote Sensing

In recent years,with the advent of the era of remote sensing big data,the method of artificial resolution alone cannot meet the need of processing massive remote sensing image data.Although the methods of object-based image analysis are very rich after decades of application research,the artificial participation of remote sensing image data classification is still very high,the development speed is also lagging behind the development of remote sensing sensors.how to quickly and effectively automatic extraction and interpreting the target information of ground objects has become a scientific problem urgently needed to be solved in the remote sensing science fields.The development of artificial intelligence and machine learning algorithms have brought new help and ideas to the classification of remote sensing images.Random Forest,as an important machine learning algorithm,has gained great attention in recent years due to its advantages of small sample size and strong stability,and its excellent performance in processing remote sensing image classification problems,which has been widely applied in object-oriented image analysis.However,in practical applications,Random Forests also has its own deficiencies.In this paper,the problems of random forest classification are analyzed,which are summarized as follows: Sensitivity to the sample imbalance;Too many redundant features are likely to cause the "Hughes phenomenon";the optimal parameters of the model need to be determined.To solve these problems,the paper studies random forest Classification optimization from sample selection optimization,feature selection optimization and model parameter optimization to improve the stochastic forest classification performance.For the problem of unbalanced samples,this paper proposes an improved SMOTE algorithm—pure-SMOTE algorithm.This method considers the distribution of surrounding samples when synthesizing samples and avoids noise generation and sample overlap.The comparative test was used to verify the validity of the pureSMOTE method.For the problem of feature redundancy,this paper introduces the SEaTH method,a commonly used feature selection method in object-oriented image analysis,into Random Forest.Experiments show that the optimized random forest not only ensures the classification accuracy but also greatly reduces the classification time.By studying the model parameters of the random forest classification method,a variable step grid search method based on empirical constraints is proposed.By using this method,the best model parameters are selected,and the classification time is greatly reduced under the condition of ensuring the classification accuracy.Using this method to select the best model parameters can greatly reduce the classification time and improve the classification efficiency of Random Forest under the condition of ensuring the classification accuracy,which lays a foundation for the application of remote sensing big data in the later period.Finally,this paper compared the optimized Random Forest with the traditional classification method and the non-optimized Random Forest through a comprehensive experiments and proved the effectiveness of the optimization.The research in this paper has certain theoretical significance and application value for high precision remote sensing information extraction...