基于深度学习方法提取苜蓿人工草地空间分布信息

Extracting Spatial Distribution Information of Alfalfa Artificial Grassland Based on Deep Learning Method

近年来,我国人工草地建设进入了快速发展阶段,及时、准确地获取大区域范围内人工草地空间分布信息,有助于提升草业宏观管理的数字化水平。以内蒙古阿鲁科尔沁旗的苜蓿人工草地为例,基于Sentinel-2多时相遥感图像,结合苜蓿光谱反射率随刈割期变化的规律,利用ENVI NET 5深度学习模型提取苜蓿人工草地空间分布信息,并与面向对象的随机森林模型提取结果进行了比较。结果表明:(1)深度学习方法提取的2020年和2021年苜蓿草地的用户者精度分别为100.00%和98.80%,制图精度分别为98.42%和100.00%,总体精度分别为99.17%和99.33%,Kappa系数分别为0.98和0.99。随机森林模型的用户者精度分别为98.90%和96.61%,制图精度分别为97.82%和94.48%,总体精度分别为98.28%和94.68%,Kappa系数分别为0.94和0.90。深度学习模型提取的苜蓿人工草地空间分布信息各项精度指标均优于随机森林模型。(2)深度学习方法提取结果中的噪声像元较少,简化了复杂的分类后处理流程。此外,深度学习模型无需逐年建模,普适性强。综上所述,深度学习模型可以精确、便利地提取研究区域苜蓿人工草地信息,具有在类似地区大面积推广应用的潜力。

The construction of artificial grassland in our country has entered a stage of rapid development in recent years.To obtain the spatial distribution information of artificial grassland in large areas timely and accurately is helpful to improve the digital level of grassland macro-management.In this study,we used Sentinel-2 multi-temporal remote sensing images of alfalfa artificial grassland in Alukorqin Banner,Inner Mongolia,combing with the variation of alfalfa spectral reflectance with cutting time,the spatial distribution information of alfalfa artificial grassland was extracted by ENVI Net-5 deep learning model,and the results were compared with the object-oriented random forest model.The results showed that:(1)The user accuracy of alfalfa grassland extracted by deep learning method in 2020 and 2021 was 100.00%and 98.80%,the mapping accuracy were 98.42%and 100.00%,the overall accuracy was 99.17%and 99.33%,and the Kappa coefficient was 0.98 and 0.99,respectively.The user accuracy of the random forest model was 98.90%and 96.61%;the mapping accuracy were 97.82%and 94.48%,the overall accuracy was 98.28%and 94.68%,and the Kappa coefficients were 0.94 and 0.90,respectively.The precision indexes extracted by the deep learning model were better than those extracted by the random forest model.(2)There were fewer noise pixels in the extraction results of deep learning,which simplified the complex classification post-processing process.In addition,the deep learning model does not need to be modeled year by year and had strong universality.In conclusion,the deep learning model can accurately and conveniently extract the information of alfalfa artificial grassland in the study area,and had the potential to be widely applied in similar area.