基于多时相Sentinel-2A影像的狼毒分布识别

Identification of Stellera chamaejasme distribution based on multi-temporal Sentinel-2A images

狼毒(Stellera chamaejasme)是近年来青藏高原高寒草地的主要入侵毒杂草之一,及时高效的调查与监测可为狼毒综合防控与退化草地恢复提供重要的技术支持。本研究选取花期前与盛花期的Sentinel-2A多光谱影像,将Google Earth Engine平台去云、环境要素掩膜、特征优选和随机森林分类相结合,探讨区域尺度的狼毒遥感识别方法。结果表明,通过狼毒敏感指数计算,以及Spearman秩相关性分析与随机森林重要性排序相结合的二次降维,提取了7项狼毒分类特征并构建了4个特征组合方案。与单时相特征组合相比,多时相特征组合有效提高了狼毒识别精度,其中,基于随机森林模型的6个特征组合方案的分类总精度为84.62%,狼毒分类精度均大于80%,识别效果最佳。本研究显示,影像去云及掩膜预处理能够有效减少分类干扰信息,花期前与盛花期提取的多时相特征组合增强了狼毒群落与其他群落的影像光谱差异,在区域尺度狼毒遥感识别中具有较好的应用潜力。

Stellera chamaejasme has been one of the main invasive noxious weeds in the alpine grasslands of the Qinghai-Tibet Plateau in recent years.Timely and efficient investigation and monitoring can provide important technical support for the integrated control of S.chamaejasme and the restoration of degraded grasslands.In this study,Sentinel-2A multi-spectral images of before flowering and full flowering were selected,and the Google Earth Engine platform cloud removal,environmental factor masking,feature selection,and random forest classification were combined to explore a regional scale remote sensing identification method for S.chamaejasme.We found that seven features of S.chamaejasme classification were extracted and four feature combinations were constructed through the calculation of the S.chamaejasme sensitivity index and the secondary dimensionality reduction,combined with Spearman rank correlation analysis and random forest importance ranking.Compared with the combination of single temporal features,the combination of multi-temporal features effectively improved the recognition accuracy of S.chamaejasme.Among them,the total classification accuracy of the six features combination scheme based on the random forest model was 84.62%,and the classification accuracy of S.chamaejasme was all more than 80%,showing the best recognition effect.This study shows that the image cloud removal and mask preprocessing can effectively reduce the classification interference information,and the combination of multi-temporal features extracted before flowering and in full bloom enhances the spectral difference between S.chamaejasme community and other plant communities,indicating good application potential in the regional scale remote sensing recognition of S.chamaejasme.