Classification Of High Resolution Remote Sensing Image Based On Deep Forest

High resolution remote sensing images are widely used in surveying and mapping,military,environmental monitoring and other fields.With the launch of high-resolution series satellites,the application of high-resolution remote sensing is gradually increasing.High resolution remote sensing image classification is the premise of remote sensing application.How to make good use of these remote sensing images with large amount of data and rich information has always been a research hotspot at home and abroad.At present,the traditional remote sensing image classification methods have low digitization efficiency,cumbersome process,long cycle and poor timeliness,so they are not suitable for the classification of large-scale remote sensing images.In addition,although the deep learning remote sensing image classification method can achieve good results,it still has some disadvantages,such as extremely dependent on label data and complex parameter adjustment and optimization of the model.In view of this,starting from the requirements of efficient and accurate classification of high-resolution remote sensing images,aiming at the disadvantages of the above classification methods,based on the deep forest model,according to the advantages and disadvantages of the deep forest model and combined with the characteristics of high-resolution remote sensing images,this paper proposes two improved deep forest models for remote sensing image feature classification: deep cascade typical correlation forest and deep cascade rotating forest.1.The influence of feature diversity on high-resolution remote sensing image classification is studied.Aiming at the fact that there are only four bands in high-resolution remote sensing image,the deep forest algorithm is used to extract and classify the features.The effect of automatic feature extraction by deep forest algorithm through remote sensing image and its application in image classification is studied.2.Aiming at the advantages and disadvantages of base classifier random forest in deep forest model,combined with remote sensing image classification,this paper improves and optimizes: expand its complexity by increasing the number of layers of deep decision forest;The new principal component analysis model is constructed by using the forest rotation classifier(PCA)data set;Based on the deep forest model,the typical correlation forest algorithm optimization is introduced to construct the deep cascade typical correlation decision forest model.Finally,the GF-2 remote sensing image data is used for classification and analysis,and the international standard data sets Indian pines data set and Pavia university data set are used to verify the feasibility of the model,and the results are analyzed and discussed.3.The problem of parameter adjustment and optimization of deep forest is studied.Combined with the characteristics that the number of labeled samples is the main factor determining the classification accuracy in high-resolution remote sensing image classification,and the feature diversity and feature utilization are important factors affecting the classification accuracy,the parameter adjustment and optimization of deep forest model is carried out.Experimental research shows that compared with CNN,random forest and deep forest,the improved deep forest algorithm is higher than the above three algorithms.In the classification of GF-2 remote sensing image,the overall classification accuracy is 98.872%,kappa coefficient is 0.9853 and F1 value is0.7641.At the same time,this method has the same excellent effect in hyperspectral remote sensing image classification.It has the advantages of low data requirements,strong robustness,few model parameters,accurate classification and high efficiency.It provides a certain reference for future scientific research.There are 22 figures,10 tables and 76 references.