摘要
水稻种植信息提取对于保障粮食安全具有重要意义。以往的研究主要是基于水稻不同生长阶段的特点进行多时相图像特征选择用于水稻制图,以安徽省寿县为研究区,提出从排除干扰地物的角度特征选择策略,基于不同物候期水稻与干扰地物之间的光谱差异,在多时相Sentinel-2图像中提取出能有效区分水稻与干扰地物的时序特征。综合利用Jeffries-Matusita(J-M)距离、随机森林(Random Forest,RF)-递归消除和皮尔森相关性分析筛选出最优指数特征集,分别采用RF、支持向量机(Support Vector Machine,SVM)和光梯度提升机(Light Gradient Boosting Machine,LGBM)3种分类算法进行水稻制图,对3种分类方法进行了精度评价,将最优分类结果与其他水稻制图产品和水稻制图方法提取结果进行比较。实验结果表明,基于筛选出的最优指数特征集,3种分类算法的水稻制图总体精度(Overall Accuracy,OA)均超过0.96,RF分类算法OA、用户精度(User Precision,UA)、面积精度最高。与其他水稻制图产品和水稻制图方法提取结果相比,所提方法能有效减少水稻错分和漏分,面积精度最高,为基于多时相影像实现精准水稻制图提供了新思路。
Rice planting information extraction is of great significance for ensuring food security.Previous research mainly focused on multi-temporal image feature selection for rice mapping based on the features of different growth stages of rice.By taking Shouxian County in Anhui Province as the study area,a feature selection strategy from the perspective of excluding interfering objects is proposed.Based on the spectral differences between rice and interfering objects at different phenological stages,temporal features that can effectively distinguish rice from interfering objects are extracted from multi-temporal Sentinel-2 images.In addition,the Jeffries-Matusita(J-M)distance,Random Forest(RF)-recursive elimination and Pearson correlation analysis are used to screen out the optimal index feature set.Three classification algorithms,RF,Support Vector Machine(SVM)and Light Gradient Boosting Machine(LGBM),are used for rice mapping.The accuracy of the three classification methods is evaluated,and the optimal classification results are compared with other rice mapping products and rice mapping method extraction results.The experimental results show that based on the selected optimal index feature set,the Overall Accuracy(OA)of rice mapping of the three classification algorithms exceeds 0.96,and the RF classification algorithm has the highest OA,User Accuracy(UA),and area accuracy.Compared with the extraction results of other rice mapping products and rice mapping methods,the proposed method can effectively reduce the commission and omission errors in rice classification while achieving the highest area accuracy,providing a novel solution for precise rice mapping based on multi-temporal images.
作者
赵萍
周俊
张树衡
吴松
常杰
申奥
ZHAO Ping;ZHOU Jun;ZHANG Shuheng;WU Song;CHANG Jie;SHEN Ao(School of Resources and Environmental Engineering,Hefei University of Technology,Hefei 230041,China;Anhui Institute of Geophysical and Geochemical Prospecting Techniques,Hefei 230041,China)
出处
《无线电工程》
2025年第6期1244-1255,共12页
Radio Engineering
基金
安徽省自然资源厅公益性地质调查项目(2021-g-2-7)。
