基于迁移学习和非监督分类的制种玉米遥感识别方法

Remote Sensing Identification for Seed Maize with Integrated Migration Learning and Unsupervised Classification

作物遥感识别主要基于监督分类方法,对样本的数量、分布要求较高,而农作物样本目视解译困难。为提高已采集样本的利用率,同时降低精细分类中对样本的依赖,本文将迁移学习与非监督分类方法相结合,在源域内构建特征工程,包括:BLUE、GREEN、RED、EDGE1、EDGE2、EDGE3、NIR、SWIR 8个原始光谱波段,以及NDVI、EVI、RVI、GNDVI、TVI、DVI、MSAVI、GCVI、RNDVI、NDRE、RRI1、RRI2、MSRRE、CLRE、IRECI、LSWI、GCI、SIPI 18个植被指数,提取出最能表征制种玉米与大田玉米冠层光谱差异,且在不同的源域内制种玉米之间差异最小的特征,将其作为先验知识用于目标域的分类任务中,再基于K-means进行制种玉米识别和制图。结果表明,在众多特征中,近红外原始波段表现出最强的优势,且在制种玉米母本去雄期后表征效果最好。计算此时间段内NIR的线性回归斜率作为特征,相较于直接基于NIR原始波段特征分类精度有所提升。利用K-means方法对2019年、2020年石河子市和奎屯市的制种玉米分类,2个目标域制种玉米2019年F1值分别为74.35%和64.97%,2020年F1值分别为72.50%和75.69%。本方法通过提取先验知识,引入非监督分类器,有效提高了样本利用率。通过提取波段回归斜率作为特征为原始波段的特征增强提供了思路,同时也为无样本场景下农作物精细分类绘图提供了方法。

Crop classification studies generally focus on different types of crops,while there are fewer studies on the fine classification of different cropping patterns of the same crop.Research on the spatial distribution of seed production is essential to control the maize market,as private seed production and concealment of acreage are occurring in the maize seed market.Seed maize and common maize are the two modes of maize cultivation.Accurate identification of both and remote sensing mapping of the spatial distribution of seed maize are essential for maize seed industry and food security.Traditional crop remote sensing classification methods require a high number and distribution of samples,while visual interpretation of crop samples is difficult.How to improve the utilization of collected samples and at the same time reduce the dependence on samples in fine classification is a pressing issue nowadays.Based on this,combining migration learning with unsupervised classification methods,firstly,using the idea of transfer learning,Linze and Wuwei were used as the source domain,and the feature engineering in the source domain was constructed,including 8 original spectral bands BLUE,GREEN,RED,EDGE1,EDGE2,EDGE3,NIR,SWIR,and 18 vegetation indices NDVI,EVI,RVI,GNDVI,TVI,DVI,MSAVI,GCVI,RNDVI,NDRE,RRI1,RRI2,MSRRE,CLRE,IRECI,LSWI,GCI,SIPI.Then,the features that best characterized the differences in canopy spectra between seed maize and common maize and that differed least between seed maize in different source domains were extracted.Finally,it was used as prior knowledge in an unsupervised classification task in the target domain.The results showed that the near-infrared primordial band exhibited the strongest advantage among the many features.By comparing the classification accuracies of three time-series ranges,namely,before the removal of male ears from the seed maize females,after the removal of male ears from the seed maize females,and during the full-life span of maize growth,the NIR bands were best characterized after the removal of male ears from the seed maize females,i.e.,when the DOY was 125~210.In order to further extract the NIR primary band information and enhance the performance of the features,the slope of the linear regression equation in the NIR band of the seed maize female after removal of male ears was used as a feature,and the K-means unsupervised method was used to classify the seed maize in the target domain of Shihezi and Kuitun.After comparative experiments on two target domains in 2019 and 2020,this method mostly showed different degrees of improvement over the K-means classification accuracies characterized by the near-infrared primitive bands in the same time period.In the two target domains Shihezi and Kuitun seed maize had F1 value of 74.35%and 64.97%in 2019 and 72.50%and 75.69%in 2020,respectively.This method effectively improved the utilization of samples by extracting prior knowledge and introducing an unsupervised classifier.It provided ideas for feature enhancement of the primary band by extracting the slope of the band regression as a feature,and provided a method for fine classification mapping of crops in sample-free scenarios.