基于 Sentinel-1 VH极化与深度学习的岳阳县早稻种植区提取

Extraction of early rice planting areas in Yueyang County based on Sentinel-1 VH polarization and deep learning

关键词:  Sentinel-1  SAR  早稻种植区  多极化  多时相  深度学习

Key words:  Sentinel-1  SAR  early rice planting area  multi-polarization  multi-temporal  deep learning

摘要:

【目的】 针对南方水稻种植区受云雨天气影响、光学遥感难以稳定获取关键生育期信息用于水稻识别的问题，以湖南省岳阳县为研究区，利用多时相Sentinel-1影像进行早稻种植区域提取研究。【方法】 文章以Sentinel-1IW模式VH极化多时相影像为核心数据源，叠加国土调查耕地矢量数据构建研究范围；基于早稻全生育期雷达后向散射时序特征构建样本集，分别采用DeepLabV3+、SegUNet和SegUNet++ 3种语义分割模型进行像素级提取，并与随机森林、支持向量机2种传统机器学习方法开展对照试验，统一采用总体精度、kappa系数等指标定量评估。【结果】 （1）在VH极化下，3种深度学习模型DeepLabV3+、SegUNet、SegUNet++早稻提取的判识精度分别为84.80%、83.10%和80.32%，显著优于随机森林（72.40%）和支持向量机（72.24%）2种传统机器学习模型。（2）DeepLabV3+模型在漏判控制、边界完整性与复杂地块适应性上最优，综合性能领先。【结论】 在多云雨区早稻提取场景中，深度学习模型凭借深层特征挖掘与空间结构学习能力，显著优于依赖人工浅层特征的传统机器学习方法，研究结果可为南方多云雨地区早稻种植面积监测与农业精细化管理提供技术参考。

Abstract:

[Purpose] The persistent cloudy and rainy weather in southern rice-growing regions makes it difficult to obtain key growth period information for rice identification using optical remote sensing. Taking Yueyang County，Hunan Province as the study area，this study investigates the extraction of early rice planting area based on multi-temporal Sentinel-1 images.[Method] Taking multi-temporal Sentinel-1 IW mode VH polarization images as the core data source，this study superimposed the cultivated land vector data of national land surveys to define the research scope，constructed a sample dataset based on the temporal backscattering characteristics of early rice throughout the whole growth period，and adopted three semantic segmentation models including DeepLabV3+，SegUNet and SegUNet++ for pixel-level early rice extraction. Random forest and support vector machine （SVM） were selected as two traditional machine learning methods for comparative experiments，and overall accuracy and kappa coefficient were uniformly adopted for quantitative model evaluation.[Result] （1） Under VH polarization，the overall accuracy of DeepLabV3+，SegUNet and SegUNet++ for early rice extraction was 84.80%，83.10% and 80.32%，respectively，significantly outperforming the two traditional machine learning models，random forest （72.40%） and SVM （72.24%）. （2） Among all models，DeepLabV3+ presented optimal performance in omission error control，boundary integrity and adaptability to complex farmland plots with the best comprehensive performance.[Conclusion] In the early rice extraction scenarios of cloudy and rainy regions，deep learning models significantly outperform traditional machine learning methods that rely on artificial shallow features by virtue of their capabilities of deep feature mining and spatial structure learning，and the research results can provide a technical reference for early rice planting area monitoring and refined agricultural management in cloudy and rainy areas of southern China.