| Hani National Nature Reserve,located in the center of Longgang Mountain at the foot of Changbai Mountain in the east of Jilin Province,is the third internationally important wetland in Jilin Province after Xianghai and Momoge.Its geomorphological types are mainly terraces composed of middle and low hills and basaltic molten rocks.This area is rich in wetland resources,such as freshwater forest swamps,and has abundant peat deposits in China,which play an important role in conserving and purifying water sources,regulating atmospheric circulation and improving the climate of small areas in protected areas.It is of great significance to carry out wetland remote sensing monitoring and accurate interpretation in time.Therefore,this paper takes Hani National Nature Reserve as the research area,integrates the advantages of Sentinel-1 multi-polarization radar and Sentinel-2 multi-spectral remote sensing images,and makes full use of 3S technology to tap the potential of Sentinel-1multi-polarization VV and VH radar band characteristics and Sentinel-2 red edge index characteristics in fine classification of forest swamp remote sensing.Using the traditional remote sensing information extraction and classification methods such as Supervised and Unsupervised Classification,Random Forest machine learning and Convolutional Neural Network deep learning,the information of freshwater forest swamp wetland cover in Hani National Nature Reserve was extracted efficiently and accurately,and the accuracy of each classification method was evaluated,and the freshwater forest swamp wetland cover in Hani National Nature Reserve in 2020 was finely classified.The main conclusions are as follows:(1)Based on the traditional Supervised and Unsupervised Classification methods,the accuracy of accurate information identification and extraction of forest swamp in Hani National Nature Reserve is relatively poor,and the classification accuracy is rough,resulting in the phenomenon of wrong identification of large-scale ground feature types and wrong assignment of similar ground feature types.The overall accuracy and Kappa coefficient of the Supervised Classification algorithm are low,which are 79.61%and 0.76,respectively.The overall accuracy and Kappa coefficient of Unsupervised Classification algorithm are 68.07%and 0.63,respectively,which is lower than that of Supervised Classification algorithm.However,with the help of traditional Supervised and Unsupervised Classification methods,remote sensing monitoring and recognition of large-area ground features still has the characteristics of high speed and low consumption.(2)The accuracy of Random Forest extraction of forest swamp wetland cover information in Hani National Nature Reserve increases at first and then tends to be stable with the increase of the number of classification trees.The optimal number of Random Forest classification trees for forest swamp information extraction in nature reserve is affected by many factors such as sample training and model parameters.The overall accuracy and Kappa coefficient of information extraction of Random Forest algorithm in Hani National Nature Reserve are 88.88%and 0.87 respectively,and the producer accuracy and user accuracy of information extraction of forest swamp are as high as 89.01%and 86.17%respectively.The research quality of identification,extraction and classification of wetland information by Random Forest algorithm is good,and it has basically met the experimental requirements..(3)The Convolution Neural Network deep learning algorithm of forest swamp landscape information in Hani National Nature Reserve has the best overall accuracy and Kappa coefficient effect,which are 93.51%and 0.92,respectively.Among them,the producer accuracy and user accuracy of forest and swamp are as high as 94.51%and 95.56%,respectively.The information recognition and extraction results of Convolution Neural Network algorithm meet the classification requirements.Compared with Random Forest,Supervised Classification and Unsupervised Classification,the overall accuracy is improved by 5.21%,17.46%and 37.37%respectively.Kappa coefficient increased by 5.75%,21.05%and 46.03%respectively.The precision of forest extraction producers increased by 6.18%,21.14%and 26.47%respectively.The accuracy of forest extraction users increased by 10.90%,19.78%and 19.45%respectively.(4)There are many types of ground objects in Hani National Nature Reserve.According to the results extracted by Convolution Neural Network algorithm,the forest swamp area is26.34 km~2,accounting for 88.75%of the natural wetland area and 10.86%of the total area of the reserve,mainly distributed in the northeast of Hani National Nature Reserve.The river covers an area of 0.71 km~2,with the smallest specific gravity,and runs through the whole area.In addition,lakes and herbaceous swamps are also important components of natural wetlands in Hani National Nature Reserve,with an area of 1.26 km~2 and 1.37 km~2 respectively.To sum up,integrating the image data of Sentinel-1 multi-polarization VH and VV radar back scattering coefficient band and Sentinel-2 red edge band,it has the advantages of Sentinel-1 synthetic aperture radar SAR in high spatial resolution,active polarization mode,penetrating clouds and vegetation cover,wide scanning range,no light interference and Sentinel-2 in monitoring vegetation growth and health status in visible light,near infrared,short-wave infrared and red edge.Using hyperspectral and high-precision fused image data and Convolution Neural Network deep learning classification algorithm,the accuracy of extracting and classifying forest swamp information in Hani National Nature Reserve is the best.Therefore,with the help of Sentinel-1 and Sentinel-2 fusion images and Convolution Neural Network deep learning algorithm,the method of wetland forest swamp information identification and extraction is worth popularizing.The research can provide data support for the corresponding departments to make scientific wetland utilization and restoration decisions. |
PDF Full Text Request