Wetlands Information Extraction Technology Study Based On Landsat8 Images For China

Objectively retrieving the quantity and distribution of wetland resources in China and establishing a complete and continuous wetland scientific data set is a major industry demand for wetland protection and restoration.The wetland ecological environment is complex and the internal accessibility is poor.The traditional wetland survey cycle is long,and the survey time and summary time of each province and city are not synchronized.At present,the most commonly used method for national scale wetland information extraction based on remote sensing is still manual visual interpretation,which is low efficient and not objective enough.In view of the above problems,this study aims to find a fast and accurate wetland information extraction method for large areas by analyzing the remote sensing characteristics of wetlands and comparing different methods,and to conduct the national wetland remote sensing classification.The results provided a data support for the wetland distribution and optimized decision-making in China.In this study,the whole nation was used as a research area,and a total of 506 scenes of Landsat8 OLI images covering the national land area in 2015 were used as the basic data source,and 224 scenes of Landsat8 and GF-1 were as supplementary data.Taking the difficulties of wetland type extraction technology,the impact of human activities,and the practical operability into account,all wetlands of the country were divided into Zone I and Zone II,which was discussed and analyzed separately.The main contents of this paper are as follows:(1)In Zone ? regions,Beijing was selected as the experimental area as it has more artificial wetlands and more human activities.The method of extracting wetland information with strong influences from human activities was discussed.Comprehensive analysis of the spectrum,shape,texture characteristics and spatial information of the image,assisting the second national wetland resource survey data,2015 land use map and other data,selecting the wetland characteristics,and studying the wetland extraction method in areas with strong human activities.The categorical sampling method-non-proportional distribution method randomly distributed verification samples were used to verify the classification results.The results show that the overall accuracy of the object-oriented CART decision tree method was 87.83%.Compared with the maximum likelihood method,the total accuracy was improved by nearly 25%.It was similar to the object-oriented nearest neighbor method,but for some wetland types,such as artificial wetlands,river wetlands,and swamp wetland,user accuracy or producer accuracy increased by about 10%.The accuracy of extracting non-wetland features by the full convolutional network method could reach about 90%,but the classification accuracy of various wetland types was low.The classification accuracy of the object-oriented CART decision tree method was obviously improved,and it has strong practicability.The multiscale segmentation algorithm assisted spectral difference segmentation technique was an effective means for extracting wetland types with strong influences on human activities and more debris.(2)In the Zone ? regions,Jiangxi province with more natural wetlands and less human activities was selected as the experimental area.The spectrum,shape,texture features and spatial information of the image were analyzed comprehensively.The characteristics of wetland extraction were selected,and the extraction methods for the areas with weaker influences on human activities and obvious differences in wetland types were studied.Accurate verification of classification results by assisting existing data and randomly distributed verification samples.The results show that the overall accuracy of the hierarchical classification method was 85.76%,which was nearly 20% higher than the accuracy of the maximum likelihood method.By calculating the different characteristic indices and then masking the non-wet ground objects,the interference of similar spectral information was reduced,and the uncertainty caused by different mechanisms of other classification methods was avoided,and the precision was improved.According to the non-zonal distribution characteristics of the wetland,the gradual extraction of the target features could greatly reduce the phenomenon of misclassification.The hierarchical classification method has a better effect on the wetland information extraction in the experimental area,and was more flexible in the wetland information extraction for large areas.(3)According to the characteristics of Zone I and Zone II,the object-oriented CART decision tree method and hierarchical classification method was applied to Zone I and Zone II,and the extraction results of five wetland types were compared with the second national wetland resource survey data.The extracted area of artificial wetland was about 4.58 million hectares.The river wetland area was about 9.09 million hectares and the lake wetland area was about 8.53 million hectares.The marsh wetland area was about 17.81 million hectares and the area of offshore and coastal wetlands was about 5.62 million hectares.The total wetland area of the country was about 45.63 million hectares.