| Terraces are the most important soil and water conservation measures and agricultural production measures in the Loess Plateau,the main region of soil and water loss and the key region of ecological environmental construction in China.The purpose of this study is to monitor the spatio-temporal change of loess terraces in a long time series efficiently and accurately by remote sensing.Google Earth Engine(GEE),a cloud-based platform of remote sensing with high-performance computing resources,was used in this study.Guyuan of Ningxia Hui Autonomous Region,a gully region of the Loess Plateau,was taken as the research area.Firstly,Guided by the idea of greedy algorithm,supervised recognition technology was used on Landsat long-term remote sensing data to optimize the recognition scenario from image features,machine learning algorithms and parameters,training sample size,and LandTrendr algorithm fitting recognition results.Secondly,confusion matrix was used as accuracy evaluation standard to construct the optimal loess terrace recognition method.Finally,based on the optimal recognition results of loess terraces,together with meteorological,topographic and regional statistical data,the variation rules and characteristics of terraces area and its fragmentation,fractional vegetation cover,spatial distribution of terraces,time to use terraces,area transition matrix,forecast of terraces area and driving factor of terraces were analyzed in research area.The main results are as follows:(1)The combination of percentage features fusion image and random forest machine learning algorithm based on sample test had the highest overall accuracy of 94.10%and Kappa coefficients of 0.87.Using more training sample size or random trees in random forest,as well as applying LandTrendr algorithm to fit the sequence of recognition results of terraces,which can improve recognition accuracy.The optimal recognition result was verified by field patch,with overall accuracy of 93.33%and Kappa coefficient of 0.80.(2)The classification distance can reflect recognition ability in machine learning,and can also be used as reference data in sampling operation to optimize sample collection strategy,thereby reducing intensity and cost of sampling work.(3)Terraces were mainly distributed on both sides of Liu-p’an Mountains in research area.From 1988 to 2019,the area of terraces in research area decreased by 45.90%,at the same time,FVC increased by 52.44%and terraces fragmentation exacerbated.In the east of research area,time to use terraces was shorter than that in the west,but the scale of terraces converted-out was greater than that in the west.However,the terrain conditions of newly-built terraces were better than that of terraces converted-out.In addition,In the next 5 years,the area of terraces in research area will still decrease.(4)The main factors affecting time to use terraces(spatial change)in research area were precipitation and maximum temperature,which can explain 94.40%of the total variation.The main factors affecting terraces area(temporal changes)were fractional vegetation cover,precipitation,and unit area output value of the tertiary industry and average temperature,which can explain 67.30%of the total variation.In summary,the optimal method of this study can efficiently and accurately monitor long-sequence and large-scale loess terraces with remote sensing.The classification distance has potential application prospects in supervised classification machine learning.In the past 32 years,the proportion of terrace agriculture in research area has declined,while the quality of that has improved,which promotes the sustainable development of ecological environment.The main reasons for returning terraces to farmland included natural factors and social factors.The main natural factors were precipitation and average temperature,and the main social factors were the rapid development of cities and the decline of village. |
PDF Full Text Request