Study On Extraction Of Wide Distribution Information Of Grassland Using Feature Transferring Machine Learning Based On Neighbor Image Overlapping

Grassland,as one of the most common types of vegetation in the global,plays a very important role in maintaining global ecological security,preventing desertification and conserving water.Accurately obtaining spatial distribution information of grassland,which not only has very important ecological and environmental significance,but also helps to formulate corresponding protection and management policies.Traditional grassland monitoring methods mainly use the field sampling method.This method is time-consuming,labor-intensive,low-efficiency,high-cost,and limited by many human factors.It is not possible to quickly perform dynamic monitoring of grassland cover information on a large scale.Remote sensing monitoring technology has developed rapidly in recent decades and has played a significant role in obtaining global land cover information.For the extraction of land cover information in large area,it is often difficult to obtain remote sensing images of the same phase,and remote sensing images of different phases are often affected by factors such as light,angle,and terrain to cause certain spectral differences,difficulty in obtaining training samples,and it is difficult to handle the image overlap area.Therefore,for a large area research area covering multiple remote sensing images,how to achieve high-precision extraction of grassland cover information remains to be explored.First,due to the geospatial consistency of the overlapping areas between adjacent images,and the types of features generally do not change significantly over time,using the overlapping area image object as an effective area for automatic selection of additional samples can effectively reduce samples selecting the workload has a positive effect on the extraction of image feature information.Second,Joint Distribution Adaptation(JDA)and Balanced Distribution Adaptation(BDA)as the transfer learning can weaken the differences between different data domains through the transformation between data features.The use of transfer learning method to balance the temporal differences between remote sensing images can weaken the change in image data features caused by temporal differences.Third,the use of machine learning classifier algorithm to complete remote sensing image information extraction can greatly improve its level of automation,the selection of a classifier algorithm with strong classification ability has a greater impact on the information extraction results.Based on the above three points,his paper proposes an automatic grassland information extraction strategy using feature transferring machine learning based on neighbor image overlapping,and achieve high-precision extraction of grassland cover information in the Qilian Mountain Nature Reserve in Gansu.In this paper,the grassland vegetation distribution in the Qilian Mountain Nature Reserve in Gansu is taken as an example,mainly based on Sentinel-2 image data and supplemented by 30 m digital elevation model data.Comparative analysis was carried out through experiments on the influence of different selection ratios of labeled samples in overlapping area on the classification accuracy,experiments on the influence of transfer learning method on the classification accuracy,comparative experiments on four classifier algorithms,Experiment of extracting grassland cover information from Qilian Mountain Nature Reserve in Gansu,and comparative experiments on five classification strategies.The following conclusions were reached:(1)In this paper,the automatic grassland information extraction strategy using feature transferring machine learning based on neighbor image overlapping can effectively and accurately extract grassland cover information from the Qilianshan National Nature Reserve in Gansu,and to a certain extent meet the thematic information of large area remote sensing images extracting demand.In this paper,1 scene image classification is used.The transfer learning model balances the image phase differences,adaptively completes the overlap region information transfer,and expands and completes the 12 scene target image classification.The average overall classification accuracy of the study area reached 92.59%,the average kappa coefficient reached 0.84,the overall classification accuracy of the single-view image reached 88.86%,and the kappa coefficient reached 0.75.The experimental results demonstrate that the method proposed in this paper can effectively realize the extraction of grassland cover information in the Qilian Mountain Nature Reserve in Gansu with few training samples.(2)The comparative experiment of grassland information extraction is performed by mixing different proportions of overlapping area labeled samples and source training samples.It can be found that with the increase of the number of labeled samples in the overlapping area,the accuracy of grassland information extraction firstly increases rapidly,then stabilizes or decreases slightly.When the proportion of labeled samples in the overlapping area reaches 10%,the average overall accuracy of grassland information extraction in the study area is the highest.(3)Through experiments on the impact of 2 transfer learning methods on classification accuracy,it can be found that the BDA model is more effective for classification than the JDA model.The BDA transfer learning method used in this paper can effectively reduce the differences in image data feature distribution caused by time differences.(4)Through comparison experiments of four classifier algorithms,in the grassland information extraction work in the study area,the BP neural network has the highest average overall accuracy and Kappa coefficient,and the strongest generalization ability.Its overall accuracy and Kappa coefficient reached 92.59% and 0.84;random forest and support vector machine followed,its overall accuracy and Kappa coefficient reached 91.61%,0.82 and 90.34% and 0.78 respectively;the generalization ability of the decision tree is the worst.Its overall accuracy and Kappa coefficient reached 87.28% and 0.74.(5)The method in this paper not only uses the overlapping area information of the image,but also adopts the transfer learning method.In order to verify the effectiveness of the two for grassland information extraction,this paper uses five classification strategies to conduct comparative experiments on grassland information extraction in the study area.The experimental results show that the fusion of overlapping image information and migration learning can improve the accuracy of grass extraction,and the fusion of overlapping area information has a greater effect on the extraction accuracy than migration learning.The use of image overlapping area information and transfer learning together is the most effective in extracting grassland information.The classification accuracy obtained by this method is slightly better than supervised classification,and it has the potential to significantly reduce the selection of training samples in the process of large area remote sensing information extraction,which can effectively improve the automation level of remote sensing topic information extraction.