基于Google Earth Engine与机器学习的黄土梯田动态监测

  目的   梯田是黄土高原最重要的水土保持措施和农业生产措施，高效、准确地获取长时间序列黄土梯田分布信息，对黄土高原的水土保持监测和评价十分重要。   方法   在Google Earth Engine(GEE)的支持下，以宁夏回族自治区固原市为研究区，使用遥感影像监督识别技术，对比随机森林(RF)、决策树(CART)、支持向量机(SVM)等3种机器学习算法的识别精度，探讨LandTrendr算法在长时间序列动态监测中的优化应用，最终获取固原市近30 a梯田分布信息。   结果   ①3种算法识别精度从大至小依次为随机森林、决策树、支持向量机。②使用随机森林算法识别梯田，基于样点检验总体精度达94.10%，Kappa系数达0.87，基于实地斑块检验总体精度达93.33%，Kappa系数达0.80。③ LandTrendr算法能有效校正时间序列中的错误值。④ 1988－2019年，固原市梯田面积减少了45.90%。⑤固原市西部的梯田使用时间较东部更长。   结论   采用本研究方法在GEE云平台可以高效、准确地遥感监测长时序、大尺度的黄土梯田。固原市近30 a梯田农业比例逐渐下降，促进了生态环境持续向好发展。图4表3参22

Dynamic monitoring of loess terraces based on Google Earth Engine and machine learning

  Objective   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 obtain the distribution information of loess terraces in a long time series efficiently and accurately, so as to monitor and evaluate soil and water loss in the Loess Plateau.   Method   Google Earth Engine (GEE), a cloud-based platform of remote sensing with high-performance computing resources, was used in this study. Guyuan City of Ningxia, a gully region of the Loess Plateau, was taken as the research area. The recognition accuracy of three machine learning algorithms, including random forest (RF), decision tree (CART) and support vector machine (SVM), was compared by using remote sensing image supervised recognition technology, and the optimized application of LandTrendr algorithm in long-time series dynamic monitoring was discussed. Finally, the distribution of terraces in Guyuan City in recent 30 years was obtained.   Result   (1) The order of identification accuracy of the three algorithms from large to small was RF, CART, and SVM. (2) Using random forest algorithm to identify terraces, the overall accuracy based on sample test was 94.10%, Kappa coefficient 0.87, and the overall accuracy based on field patch test was 93.33%, Kappa coefficient 0.80. (3) LandTrendr algorithm can effectively correct the errors in the time series and improve the accuracy of time series identification. (4) From 1988 to 2019, the area of terraces in Guyuan decreased by 45.90%. (5) The time to use terraces in the west of Guyuan was longer than that in the east.   Conclusion   The RF machine learning algorithm combined with LandTrendr algorithm on GEE can efficiently and accurately monitor long-term and large-scale loess terraces. In the past 30 years, the proportion of terrace agriculture in Guyuan City has gradually declined, which promotes the sustainable development of ecological environment. Ch, 4 fig. 3 tab. 22 ref.]