The Extraction And Analysis Of Land Cover Type Based On Multi-feature Fusion Support Vector Machine

Under the background of intensified urbanization and co-construction of ecological and environmental civilization,it is very important to obtain the information about the current situation and change of land cover type.In recent years,with the further development of artificial intelligence,more and more scholars have applied machine learning algorithms to the extraction of feature types.However,the extraction based on a single feature inevitably has the phenomenon of "different spectrum with the same matter" and "different matter with the same spectrum",and there are inextricable links between features,the analysis of time series changes based on a single feature is not enough.Therefore,considering the interdependent relationship between ground objects,taking Tongzhou district of Beijing as the research object and using Landsat images as experimental data,this paper is devoted to exploring the extraction of land cover type based on multi-feature fusion support vector machine and comprehensive analysis based on the classification results,and has achieved the following results:(1)The Gaussian difference method was used for the first time to extract the image texture features.In theory,the advantages and disadvantages of different texture extraction methods were analyzed.In order to avoid the data redundancy caused by the correlation between the bands,the OIF(Best Index Method)was used to select the 4,5,and 7 combination bands for texture extraction,and through the conventional methods such as Gray level co-Occurrence matrix and Gabor filter were compared to verify the feasibility of Gaussian difference method.The final classification effect and multi-faceted accuracy evaluation reflected that the Gaussian difference method had certain applicability in image texture extraction.(2)Three classification models based on multi-feature fusion support vector machine were constructed and comprehensively evaluated.Using the confusion matrix to quantitatively evaluate the classification results,such as overall accuracy and Kappa coefficient,combined with the land cover results map and field survey data of the 2016 BJ-2 image in the study area,qualitative evaluation was carried out,and the execution efficiency of the three models was recorded.Considering all aspects,the multi-feature fusion support vector machine model composed of Gaussian difference method and normalized difference index had the highest classification accuracy,and the multi-feature fusion support vector machine model was used to complete the land cover type classification of 6 images from 1990 to 2016.(3)The data information extracted by the multi-feature fusion support vector machine model was comprehensively analyzed,and the dynamic change process of different feature types was fitted to predict the future fluctuation trend.Based on the pixel scale statistics,the land occupation of each object was calculated.Combined with the characteristics of hysteresis,one-way variation and final balance,Three preferred mathematical models were selected,Boltzmann,Logistic and Langevin,respectively,to complete the fitting of the dynamic changes of the extracted features.According to the principle of maximum correlation and minimum error,the optimal model was selected as Boltzmann model,and the trend of the overall feature coverage type during 2016-2025 was predicted based on this model.