| 摘要: |
| 【目的】采用遥感影像对不同种类农作物耕地进行高效精准的分类，以期获取最优的农作物种植信息提取方案，为农业生产提质增效提供决策支持。【方法】以陕西省宝鸡市扶风县为研究区，利用Google Earth Engine（GEE）平台，基于2020年10月至2021年12月的Sentinel-2影像和目视解译与野外定位相结合选取的样本点，使用随机森林（random forest,RF）算法，结合归一化植被指数（normalized difference vegetation index，NDVI）、增强植被指数（enhanced vegetation index,EVI）、物候参数特征、遥感影像红边波段，且将根据NDVI计算出的物候参数的波段记为N，根据EVI计算出的物候参数的波段记为E，引入红边波段记为1，未引入红边波段记为2，在此基础上构建了8种不同的分类模型(NDVI-N1、EVI-E1、N1、E1、NDVI-N2、EVI-E2、N2、E2)，并与4种样本分割比例（训练样本数与检验样本数的比例分别为5∶5，6∶4，7∶3和8∶2）组合，共计32种分类方案，利用这些方案对扶风县的不同种类农作物耕地进行分类，并计算了不同分类方案的总体精度和Kappa系数。【结果】（1）经S-G滤波法处理后，6种地物类型的NDVI和EVI时序曲线噪声减小且更为平滑，且NDVI和EVI曲线差异明显。（2）在32种分类方案中，只有样本分割比例为7∶3时的EVI-E1模型、EVI-E2模型和样本分割比例为6∶4时E2模型的总体精度高于90%，说明在扶风地区用EVI分类效果更好一些。（3）在样本分割比例为7∶3的情况下，EVI-E1模型的总体精度和Kappa系数均最高，分别为91.66 %和89.41%，为最优分类方案。但是在该方案中引入红边波段后，其总体精度下降了1.18%，Kappa系数下降了1.53%，可知运用该分类方案时应剔除红边波段。（4）运用最优分类方案对小麦地、玉米地、土豆地的提取结果与实际调查情况大致相同。【结论】基于GEE平台和Sentinel-2数据实现了对县域级农作物种植面积的精确分类，可为扶风县农业生产提供决策支持。 |
| 关键词: 农作物分类 Sentinel-2 随机森林算法 Google Earth Engine 扶风县 |
| DOI：DOI:10.13207/j.cnki.jnwafu.2024.11.014 |
| 分类号: |
| 基金项目:国家自然科学基金面上项目（42177344）；国家自然科学基金黄河水科学研究联合基金项目（U2243213） |
|
| Farmland classification based on Google Earth Engine and Sentinel-2 data |
|
SONG Zhaoyang1, SHI Shangyu1,2, WANG Fei1,2, ZHAO Yulong2, LIU Yuanhao1
|
|
1.State Key Laboratory of Erosion and Dryland Agriculture on the Loess Plateaus,Institute of Soil and Water Conservation,Northwest A&F University,Yangling,Shaanxi 712100,China;2.Institute of Soil and Water Conservation,Chinese Academy of Sciences and Ministry of Water Resources,Yangling,Shaanxi 712100,China
|
| Abstract: |
| 【Objective】This study utilized remote sensing imagery to classify different croplands accurately to obtain optimal information extraction method and provide basis for improving the quality and efficiency of agricultural production.【Method】This study focused on Fufeng County,Baoji City,Shaanxi Province.Using the Google Earth Engine (GEE) platform,sample points were selected by visual interpretation and field positioning using Sentinel-2 imagery from October 2020 to December 2021.The random forest (RF) algorithm was incorporated with normalized difference vegetation index (NDVI),enhanced vegetation index (EVI),phenological parameters and red-edge bands of remote sensing imagery.The bands calculated based on NDVI were denoted as N,those calculated based on EVI were denoted as E,the inclusion of red edge bands was denoted as 1,and the exclusion was denoted as 2.Based on these parameters,eight different classification models were constructed (NDVI-N1,EVI E1,N1,E1,NDVI-N2,EVI E2,N2 and E2).Four sample segmentation schemes were combined with training to validation ratios of 5∶5,6∶4,7∶3 and 8∶2,resulting in a total of 32 classification schemes.These schemes were used to classify different types of croplands in Fufeng,and the overall accuracy (Ao) and Kappa coefficient of each scheme were calculated.【Result】(1) After processing with the S-G filtering method,the noises in the NDVI and EVI time series curves of six land cover types were reduced and the curves became smoother.Additionally,the NDVI and EVI curves of the six land cover types showed clear differences.(2) Among 32 classification schemes,only EVI-E1 model and EVI-E2 model with a sample split ratio of 7∶3 and E2 model with a sample split ratio of 6∶4 had overall accuracies of higher than 90%,indicating that EVI based classification was better in Fufeng.(3) With a sample split ratio of 7∶3,the EVI-E1 model had the highest overall accuracy and Kappa coefficient of 91.66% and 89.41%,respectively,making it the optimal classification scheme.However,introducing red-edge band decreased overall accuracy and Kappa coefficient by 1.18% and 1.53%,suggesting that the red edge band should be excluded.(4) The extraction results for wheat,corn and potato fields using the optimal classification scheme were roughly consistent with actual survey results.【Conclusion】This study achieved accurate classification of croplands in Fufeng based on GEE platform and Sentinel-2 data and provided valuable decision support for agricultural production. |
| Key words: classification of croplands Sentinel-2 random forest algorithm Google Earth Engine Fufeng County |
