基于Softmax分类器的小春作物种植空间信息提取

Spatial Information Extraction of Spring Crops Based on Softmax Classifier

[目的]使用浅层机器学习分类方法和多光谱遥感影像快速准确提取研究区小春作物(油菜、小麦)种植空间信息。[方法]选择研究区小春作物识别最佳时期的Sentinel 2A MSI多光谱影像,融合得到10 m分辨率影像,然后降尺度生成15、20、30 m分辨率影像,结合地面调查数据,建立油菜、小麦、林地、居民地、水体等典型地物感兴趣区,训练Softmax分类器,基于不同空间分辨率影像提取油菜、小麦种植空间信息。[结果]①基于Softmax分类器和10 m分辨率融合影像的小春作物分类总体精度为90.02%,Kappa系数为0.8344,其中油菜生产者精度和用户精度分别为93.14%、91.42%,小麦的分别为87.93%,98.09%;②Softmax法的小春作物分类精度随影像空间分辨率下降而降低,15、20、30 m分辨率影像的分类精度较10 m的分别下降9.80%、12.04%和13.04%,Kappa系数依次减少0.1538,0.1873和0.2088;③15、20、30 m分辨率影像的油菜分类精度较小麦的低,影响因素为油菜花期和种植地块破碎分散。[结论]Softmax分类器在10~30 m中高分辨率影像小春作物分类中具备较高的精度,可作为常规方法应用于业务化的作物监测工作。

[Objective]Using shallow machine learning classification methods and multi-spectral remote sensing images extracted spatial information of planting spring crops(rape,wheat)in the study area quickly and accurately.[Method]The sentinel 2 A MSI multi-spectral images of the best period of crop identification in the study area were selected.The 10 meter resolution image was obtained by a series of pre-processing such as fusion and clipping.Then 15,20 and 30 meter resolution images were generated by the way of downscaling.Typical areas of interest such as rapeseed,wheat,woodland,residential areas,and water bodies were established by combining ground survey data.The Softmax classifier was trained.Finally,the trained classifier was used to extract the spatial information of rapeseed and wheat based on different spatial resolution images.[Result]Through verification,the following results were obtained:(i)The overall accuracy of spring crop classification based on Softmax classifier and 10 meter resolution fusion image was 90.02%.The Kappa coefficient was 0.8344.The producer accuracy and the user accuracy of rapeseed was 93.14% and 91.42%,respectively.The wheat was 87.93% and 98.09% respectively.(ii)As the spatial resolution of the image decreased,the accuracy of the Softmax method for extracting spring crops also declined.Taking the10 meter spatial resolution image as a reference,the classification accuracy of the 15,20,30 meter resolution images decreased by 9.80%,12.04% and 13.04%,respectively.The Kappa coefficient also decreased by 0.1538,0.1873 and 0.2088.(iii)The classification accuracy of rapeseed at 15.,20,30 meter resolution was lower than that of wheat.The influencing factors were the rape flowering period and the planting plots broken and dispersed.[Conclusion]The Softmax classifier has high precision in the classification of spring crops with medium high resolution images of 10-30 meter,which can be used as a conventional method for operational crop monitoring.