| Tea is one of China’s traditional agricultural products,and Yunnan is one of the important tea production areas.Yunnan is recognized as the birthplace of tea trees and is also a major province for tea production in China.For many years,Yunnan’s tea planting area and comprehensive output value have ranked among the top in the country.The traditional method of obtaining the distribution area of tea gardens is mainly through on-site investigation by grassroots technical personnel,which is timeconsuming and labor-intensive.Remote sensing technology can quickly and accurately extract the planting areas of crops by processing high-resolution images taken by satellites or airplanes,realizing refined management,and helping to monitor and warn natural disasters,providing strong support for agricultural production.Therefore,based on remote sensing technology,the research on tea garden extraction is of great practical and scientific significance.This poses a major obstacle to remote sensing-based extraction of tea plantations as tea trees are usually planted in areas with large undulating terrain.The tea plantations in Yunnan are usually located in tropical and subtropical regions,with fragmented land masses and complex vegetation structures,resulting in tea plantations being easily confused with other features in the imagery.Therefore,to address these problems,this paper uses an object-oriented approach to extract tea plantations in Yunnan based on the high resolution images of HIS-2,and selects source domains and samples with good classification effects to investigate the migration effects on images of different phases using the migration learning method.The main research elements are as follows:(1)The object-oriented classification method was used to extract the tea gardens,and the image was segmented by determining the segmentation parameters through multiple tests.Support vector machine and random forest classification methods were used to explore the extraction and classification effects of different classification algorithms on Gaofen-2 remote sensing images.The recursive feature elimination method was used to optimize the classification feature space,and then the random forest classification was carried out.The results showed that the highest accuracy was obtained by optimising the feature space for random forest classification,with an average improvement of 2.43% in overall accuracy after optimisation.(2)According to the calculated feature importance,the shape feature has been eliminated 24 times,and the initial feature space of all images has a total of 36 shape features.In classification studies,the shape feature of the object contributes the least to the extraction.In the object-oriented image classification,the band standard deviation features of the object are eliminated more,and the contribution of the spectral features is greater,and the number of times of elimination is 12,and the number of elimination is the least.(3)Based on the object-oriented classification method,calculate the maximum mean difference between the image with good classification results and other images,select the image with the largest mean difference and the smallest difference as the source domain image for migration,and use the migration component analysis and joint distribution adaptation method Migration of source domain samples is carried out,and the results show that the migration classification effect of the joint distribution adaptation method is better,and the overall accuracy is 3.79% higher than the result of the migration component analysis on average.The target image with a large temporal difference from the source domain image is not as good as the migration classification effect of the target image with a small temporal difference.After migration,the overall accuracy is the highest at 91.4%,and the Kappa coefficient is 0.83. |
PDF Full Text Request