A Study Of Grassland Change Monitoring Based On Deep Learning

Grassland ecosystem is an important part of natural ecosystem,which has important value for maintaining ecological balance,regional economy and human history.Xinjiang,located in the northwest border of China,is one of the"five pastoral areas"in China.It is mainly due to its rich natural grassland resources,of which the grassland area is 5.6×10~7 hm~2,accounting for 14.6%of the total grassland area in China,ranking the third in China,with high eco-economic value.On the other hand,Xinjiang is an arid and semi-arid region,with widespread deserts and fragile ecological environment.The decline of regional grassland vegetation often indicates the beginning of desertification.Therefore,the surface vegetation and underground roots of grassland have the functions of wind and sand fixation and soil and water conservation.Therefore,grassland plays a great role in maintaining and improving the fragile ecological environment in Xinjiang,and it is very important for the real-time dynamic monitoring.Remote sensing has been widely used in grassland monitoring because of its large monitoring range and strong periodicity.However,with the development of fine animal husbandry management and ecological environment civilization,decision-making requires more and more monitoring of grassland growth.In the face of more and more multi-source and multi-resolution remote sensing data,how to quickly and accurately identify vegetation growth,especially the dynamic changes of sparse grassland,has a very practical significance.For the rapid processing and recognition of grasslands supported by remote sensing big data,this paper adopts deep learning with feature autonomous learning-capsule network algorithm,using Landsat 5,Landsat 7 and Landsat 8 remote sensing images as data sources to the Tangbula grassland.The coverage of the grassland in the test area and the changes in different periods were effectively discriminated,and good results were obtained,and the effectiveness of the capsule network algorithm was verified.The main research contents and results are as follows:(1)In this paper,based on the capsule network,grassland extraction and dynamic monitoring are realized by different improved methods.For the problem of grassland extraction,the convolution layer of capsule network is changed to the linear connection layer,combining with the characteristics of manual extraction to achieve accurate grassland extraction.Experiments show that the accuracy of grassland extraction in2009,2014 and 2019 is more than 90%.For grassland coverage monitoring,an enhanced vegetation index time series is constructed.By improving the two-dimensional convolution layer of capsule network to one-dimensional convolution layer to adapt to the feature extraction of time series,the grass coverage monitoring is realized.The experiment shows that the classification accuracy of different grassland coverage is more than 90%.(1)Aiming at the problem of easy mixing of forest and grass,the terrain and texture features of the study area are extracted,and the spectral band is combined to construct a sample suitable for forest and grass classification.By improving the feature extraction layer of the capsule network to make it suitable for one-dimensional data,the remote sensing images of the study area in different years are interpreted.The results of the study show that the grassland area in the study area decreased by 8.8%from 2009 to2014,and grassland was effectively managed from 2014 to 2019.In 2019,the grassland area increased by 0.23%compared with 2014.(2)To solve the problem of different grassland coverage monitoring,this paper uses the time series of enhanced vegetation index to build classification samples.For the time series samples,one-dimensional convolution layer is used as the feature extraction layer to classify the time series and realize the monitoring of different grassland coverage.The results showed that the area of grassland with different coverage in the study area decreased significantly in 2009-2014,especially the area with high coverage,more than 50%of which had different degrees of degradation.The area with high coverage only accounted for 26.05%of the grassland area in 2014.During2014-2019,the quality of grassland improved,and more than 50%of the grassland with medium coverage changed into high coverage grassland,with high coverage Grassland accounts for more than 40%.(3)Through the integration of the classification results of the two models,the classification results of 2009,2014 and 2019 in the study area are obtained.Through the comparative analysis of the classification results,the grassland change results are obtained.The research shows that the improper management of grassland in 2009-2014has resulted in a large area reduction of grassland,and the quality of grassland has declined seriously.The area with high grassland coverage has decreased significantly,from 49.98%in 2009 to 26.05%in 2014.In 2014-2019,this situation improved significantly,the grassland area increased slightly,with an increase of 0.23%,and a large number of medium coverage grassland changed into high coverage grassland,the overall quality of grassland increased,and the area of high coverage grassland again exceeded 40%.The overall classification accuracy of the grassland extraction model and grassland-coverage monitoring model implemented in this paper and the extraction accuracy of each category are above 90%,which can meet the requirements of comparing and obtaining change detection results.By comparing the classification results of remote sensing images in the study area,the characteristics of the temporal and spatial changes of grassland in the study area from 2009 to 2014 were grasped,which provided a practical method for grassland management.