| Land cover data acquisition and its spatial and temporal change analysis are important tools for monitoring various natural resources and ecological environments,grasping the distribution of regional vegetation types and studying the transformation of land cover types.The Qilian Mountain National Park has complex topography,large altitude differences,and obvious vertical zone differentiation patterns.It is significant to study the accuracy of the cloud computing platform and optimal machine learning algorithms for long time series land cover classification in mountainous areas,and to analyze the spatial and temporal changes of different land cover types for natural resource management and high-quality sustainable development of the Qilian Mountain National park.In this study,based on Google Earth Engine cloud computing platform,used Qilian Mountain National Park as the study area,used Landsat 5,Landsat 8,Sentinel-1,Sentinel-2 images with low cloudiness for many years and field sample data.Combining topography,spectrum,texture,tassel cap transformation and phenology features,we compared the land cover classification accuracy of four machine learning methods and two resolutions of Setinel-2 and Landsat 8 data.Finally,the random forest algorithm was selected to complete the land cover classification of Qilian Mountain National Park in 1990,2000,2010 and 2020.And the land cover changes over three decades were analyzed using the post-classification comparison method.The main results are as follows:（1）Results of land cover classification in Qilian Mountain National Park: the overall accuracies of random forest,support vector machine,classification and regression tree,and gradient boosting tree are 89.7%,78.0%,81.4%,and 90.1%,respectively.The random forest and gradient boosting tree show better classification accuracy.It was found that the random forest classification algorithm is more accurate in portraying the boundaries and wholeness of the land cover types after comparing the classification effects of these two,and its parameter adjustment range is more reasonable as a classifier.In addition,the higher resolution Sentinel-2 data had better classification accuracy,better results,and more accurate portrayal of land cover type details than Landsat 8 data,which was more in line with the actual.（2）Optimization of classification features: the combination of multiple features can improve the accuracy of land cover classification,but at the same time,too many features can have redundancy and reduce the classification accuracy.In this study,the feature set was optimized using J-M distance to reduce 37 features to 24 features for land cover classification from Sentinel-2 data and increase the overall accuracy from90.3% to 91.6%,and 35 features to 29 features from Landsat data and increase the overall accuracy from 88.6% to 89.7%.And the J-M distance indicated that the phenological features extracted from the vegetation index time series data showed good differentiability in the easily confounded vegetation categories.Second,the elevation in the topographic features also showed good differentiation in the classification of mountainous areas.（3）The sample migration based on the spectral angular distance can well achieve long time sample acquisition over a large surface area without much impact on accuracy.By migrating the 2020 sample to 2010,2000,and 1990,the total sample size decreased by 10.1%,13.7%,and 18.0%,respectively,while the classification accuracy decreased by only 2.0%,4.2%,and 6.6%.The overall accuracy reached 88.0%,86.0%,and 83.9%,respectively.It is shown that this method can achieve the acquisition of large area time series samples,while the accuracy will not be affected too much.The tedious process of visual interpretation and fieldwork in sample collection of different years is avoided,and the classification efficiency is improved.（4）Analysis of spatial and temporal distribution and changes of land cover in Qilian Mountain National Park during 1990-2020: overall,forests and scrub are mainly distributed in Gansu Province,located in Sunan County and Tianzhu County,while grasslands are abundantly distributed in both Gansu and Qinghai Provinces,mainly located in Sunan County,Qilian County and Tianjun County.From 1990 to 2020 the forest decreases by 62.2 km2,the scrub increases by 440.5 km2 and the grassland decreases by 794.7 km2.Among them,the area of forest decreased by 84.0 km2 from1990 to 2000,mainly occurred in Sunan County and Tianzhu County,the area of forest increased by 16.7 km2 from 2000 to 2010,mainly located in Minle County and Sunan County,and the area of forest increased by 5.1 km2 from 2010 to 2020,the increased area was located in Tianzhu County and Menyuan County.The area of scrub increased more from 1990 to 2000,and was basically stable from 2000 to 2020.The area of grassland,on the other hand,showed fluctuations,with a total area of 12,858.5 km2 and 13,690.2 km2 in 1990 and 2010,respectively,and a total area of 12,363.0 km2 and 12,063.8 km2 in 2000 and 2020,respectively.The vegetation change is mainly the interconversion between forest,scrub and grassland,and the change areas are concentrated in Sunan County,Tianzhu County and Menyuan County.It can be found that with the vigorous implementation of ecological and environmental protection measures in Qilian Mountain National Park,the area of the forest has gradually increased since 2000.The results of this study provided a reference for the key areas that should be paid attention to in the process of vegetation conservation,and helpd to allocate management resources and develop conservation programs in a reasonable manner. |
PDF Full Text Request