基于Sentinel数据与多特征学习的大豆种植面积提取
Soybean planting area extraction based on Sentinel data and multi feature learning
段承君 1杜晓初 2龙慧灵 2梅新 2杨贵军 3张有智4
作者信息
- 1. 湖北大学资源环境学院,武汉 430062;农业农村部农业遥感机理与定量遥感重点实验室/北京市农林科学院信息技术研究中心,北京 100097
- 2. 湖北大学资源环境学院,武汉 430062
- 3. 农业农村部农业遥感机理与定量遥感重点实验室/北京市农林科学院信息技术研究中心,北京 100097
- 4. 黑龙江省农业科学院,哈尔滨 150086
- 折叠
摘要
[目的]大豆是中国重要的粮油兼用作物,是优质蛋白的主要来源.及时准确地获取大豆的种植面积,对评估大豆振兴计划实施效果及后续政策的制定具有重要意义.[方法]文章以黑龙江省黑河市为研究区,基于Google Earth Engine平台,利用多生育期Sentinel-2A、SRTM(Shuttle Radar Topography Mission)等数据,构建经特征优选后的光谱特征、植被指数特征、纹理特征和地形特征等多特征数据集,并对比分析随机森林(Random Forest,RF)、分类与回归树(Classification and Regression Tree,CART)、支持向量机(Support Vector Machine,SVM)等算法,选择效果最佳算法对2019-2021年黑河市大豆种植面积进行提取分析,实现大豆种植面积的区域制图.[结果](1)经特征优选后,共11个特征对大豆、水稻、玉米有较好的区分度.(2)对比不同生育期遥感影像进行面积提取的结果,鼓粒期效果最佳.(3)对比不同的特征学习方法,随机森林算法在大豆鼓粒期的面积提取结果最佳.(4)以县级乡镇区划为最小统计单元,2019-2021年黑河市大豆种植面积分别为93.67万hm2、159.62万hm2、133.54万hm2.[结论]大豆种植面积先增加后减少,种植空间分布从分散的大田种植转向集中.基于Sentinel数据与多特征学习的方法能够快速、准确地提取大豆的种植面积,有助于准确掌握大豆的种植情况,为大豆的种植与管理提供依据.
Abstract
[Purpose]Soybean is an important dual-use crop for grain and oil in China and a major source of high-quality protein.Timely and accurate access to the planting area of soybean is of great significance in assessing the implementation effect of the Soybean Revitalisation Plan and the formulation of subsequent policies.[Method]In this study,Heihe City in Heilongjiang Province was taken as the study area,and based on the Google Earth Engine platform,multi-feature datasets such as spectral features,vegetation index features,texture features and topographic features after feature screening were constructed using data such as multi-fertility Sentinel-2A,SRTM(Shuttle Radar Topography Mission).Also,the algorithms such as Random Forest(RF),Classification and Regression Tree(CART)and Support Vector Machine(SVM)were compared and analyzed.The best algorithm was selected to extract and analyze the soybean planting area in Heihe City from 2019 to 2021,and to realize the regional mapping of soybean planting area.[Result]The results showed that nine features have good differentiation for soybean,rice and corn.Comparing the results of area extraction from remote sensing images of different fertility periods,the best results were obtained in the bulging grain period.Comparing different feature learning methods,the random forest algorithm has the best area extraction results in the soybean bulging stage.Taking the county township division as the smallest statistical unit,the soybean planting area in Heihe City from 2019 to 2021 was 936 700 hm2,1 596 200 hm2 and 1 335 400 hm2 respectively.[Conclusion]Soybean planting area increases and then decreases,the spatial distribution of planting shifts from decentralized field planting to concentration.The method based on Sentinel data with multi-feature learning can quickly and accurately extract the planting area of soybean,which helps to accurately grasp the planting situation of soybean and provides a basis for the planting and management of soybean.
关键词
大豆/种植面积/机器学习/Google/Earth/EngineKey words
soybeans/planting area/machine learning/Google Earth Engine引用本文复制引用
基金项目
国家重点研发计划(2022YFD2001103)
出版年
2024
国家科技期刊平台
NSTL
万方数据