节点文献
多源遥感数据特征优选的大兴安岭沿麓不同农作物提取
Multi-source remote sensing data feature optimization for different crop extraction in Daxing′Anling along the foothills
【摘要】 【目的】以内蒙古呼伦贝尔市为研究区,采用多源遥感数据进行农作物提取,旨在为多云多雨条件下、跨气候带区域和高寒旱作农业区大尺度农作物提取应用提供解决方案。【方法】基于Google Earth Engine平台,采用以Sentinel-1和Sentinel-2为主的多源遥感数据进行农作物提取。利用Boxcar滤波后Sentinel-1多极化数据构建时序数据集;获取Sentinel-2 TOA生长季数据,合成为最小云量影像并计算多种遥感指数特征;将地形因素和地表覆盖作为分类辅助数据,结合地面样本数据应用到随机森林分类器中,获取10 m精度的作物分布图,通过评估分类精度和分类贡献水平,将特征优选后的数据集用以提取研究区农作物分布,并分析农作物的分布特征。【结果】通过将雷达数据和光学数据进行组合(VV_B+VH_B+M+I+T+L)得到87.41%的农作物分类精度,比单独的组合(VV_B+VH_B、M+I+T+L)提取结果分别提高了7.58%和7.88%;使用500树的分类精度最高,当决策树的数量超过或不足500时,分类精度都有所下降,过多的决策树并没有提高分类器的性能;利用多源遥感数据特征波段重要性排名前24位组成的数据集提取的农作物总体精度为87.76%,Kappa系数为0.86。【结论】该研究结果获得了比较理想的农作物分布图,可实现常规业务化监测,能够为农作物遥感调查提供有效手段。
【Abstract】 【Objective】Using multi-source remote sensing data for crop extraction in Hulun Buir of Inner Mongolia,it aimed to provide solutions for large-scale crop extraction applications under cloudy and rainy conditions,across climatic regions and alpine dry farming areas.【Methods】Based on the Google Earth Engine platform,and used Sentinel-1 and Sentinel-2 as the main source of multi-source remote sensing data. Using Boxcar filtered Sentinel-1 multi-polarization data to construct a time series data set,and acquire Sentinel-2 TOA growth season data and synthesized it into a minimum cloud cover image and calculate multiple remote sensing index features;using topographic factors and land cover as auxiliary data for classification,combined ground sample data and apply it to a random forest classifier to obtain a 10 m precision crop distribution map. By evaluating the classification accuracy and classification contribution level,the data set after feature optimization was used to extract the crop distribution in the study area,and analyze the distribution characteristics of crops.【Results】By combining radar data and optical data(VV_B+VH_B+M+I+T+L),87.41% of crop classification accuracy was obtained,which was higher than the extraction results of a separate combination(VV_B+VH_B,M +I +T +L)7.58% and 7.88%;500 trees classification had the highest accuracy,once the number of decision trees exceeds or less than 500,the classification accuracy decreased,and excessive decision trees did not improve the performance of the classifier;the overall accuracy of crops extracted from the data set composed of the top 24 important bands of multi-source remote sensing data was 87.76%,and the Kappa coefficient was 0.86. 【Conclusion】The results of this study had obtained a relatively ideal crop distribution map,which could realize routine operational monitoring and provide effective means for crop remote sensing surveys.
【Key words】 Multi-source remote sensing data; Dry farming area; Crop mapping; Machine Learning;
- 【文献出处】 北方农业学报 ,Journal of Northern Agriculture , 编辑部邮箱 ,2020年06期
- 【分类号】S127
- 【下载频次】195