| Check dam are an effective soil and water conservation measure that has been widely constructed around the world.In the past 70 years,a large number of check dams have been built on the Loess Plateau to intercept sediment and form valuable silted field resource.Accurately obtaining the spatial location of check dam and the scope and area of silted field on the Loess Plateau is an important prerequisite for the scientific management of check dam and the evaluation of its water and soil conservation benefits.For this reason,relevant official agencies have organized large-scale surveying and mapping of check dam information for many times,which consumed a huge amount of manpower and material resources.In recent years,the popularization of high-resolution remote sensing images has provided a convenient way for the information collection of check dams.However,the mainstream method is still in the stage of visual interpretation,and several automatic identification methods have shortcomings and defects,which make it difficult to be stably extended to large areas.This study used multi-source and multi-temporal remote sensing images and terrain data to collect high-quality samples and construct a biased feature space.By applying ensemble learning classification methods based on the imbalanced classification theory,a large-scale recognition research of silted fields and dam bodies at a pixel level was conducted for check dams in the Wuding River catchment of the Loess Plateau.The identification results of silted fields were then used to quantitatively evaluate the sediment retention and carbon sequestration capacity of check dams at the basin scale.The main research contents and results are as follows:(1)The samples of non-silted field and silted filed collected based on Google Earth images are highly unbalanced in quantity,and the imbalance ratios differ significantly among basins.Since the check dam are built at the bottom of gullies,the silted filed are relatively flat and have obvious changes in land cover.The 23 classification features extracted based on two phases of Sentinel-2 images and AW3D30 DEM show good numerical and spatial consistency on the samples of silted filed.(2)Based on the data resampling strategy,under the premise of ensuring the relative balance between the producer’s accuracy and the user’s accuracy of the silted filed,three ensemble learning classification algorithms:Random Forest,Extreme Gradient Boosting and Easy Ensemble were adopted to construct the silted filed recognition model,and the recognition accuracy(F1-score)of silted filed on the test set reaches 0.7501,0.7664 and 0.7754 respectively.The hybrid ensemble model Easy Ensemble based on the Extreme Gradient Boosting classifier has the best classification performance and robustness,and can identify silted filed most accurately at the basin scale.The feature importance analysis shows that the appropriate multi-temporal spectral indices and macro-topographic features can make important contributions to the accurate identification of silted filed,especially the NDVI difference before and after the silted filed is covered by crops.(3)The Easy Ensemble model was applied to the entire Wuding River catchment to obtain the silted filed map with a resolution of 10-meter.According to the location characteristics of the check dam,the dam body of check dams was identified by combining the silted filed identification results and the flow accumulation area calculated from the hydrological analysis.The results show that by 2021,277.41 km~2 of silted filed have been preserved in the Wuding River catchment,the number of check dams exceeds 10500,and more than 90%of the check dams have been successfully identified.According to the model between silted filed area and sediment retention index and the organic carbon content of silted filed,it is estimated that the check dams in the Wuding River catchment currently hold a total of 2840 million tons of sediment,and the amount of carbon buried is about 6.248 million tons. |
PDF Full Text Request