基于多时相合成孔径雷达数据的水稻种植面积监测

Rice planting area monitoring based on multi-temporal synthetic aperture radar(SAR)data

与光学遥感相比,合成孔径雷达(SAR)遥感能够不受云雨天气影响,为大范围作物种植信息的精准监测提供新手段。本研究以天津市小站稻为例,基于2018-2021年的多时相Sentinel-1A SAR影像,提出了结合小站稻生长特征相似性分析与随机森林分类的水稻种植分布和面积监测方法。首先提取VV和VH极化方式下不同地物的后向散射系数时间序列特征曲线,并利用HANTS滤波来消除噪声影响。然后根据野外调查数据获取小站稻参考生长曲线,构建小站稻相似性指数,筛选出小站稻可能种植区域。最后采用随机森林分类模型提取小站稻种植面积。结果表明,基于多时相Sentinel-1A SAR影像相似性分析及随机森林分类能够获得较高精度的水稻种植面积,VV和VH两种极化方式下提取的水稻种植面积与统计年鉴结果的平均相对误差分别为2.67%和3.80%,总体分类精度分别达到95.52%和93.40%,Kappa系数分别为0.94和0.93;与不引入相似性指数进行分类相比,VV和VH极化方式下引入相似性指数后总体分类精度分别提高4.35个百分点和3.13个百分点,Kappa系数分别提高0.04和0.03,水稻的制图精度分别提高3.38个百分点和3.25个百分点。本研究结果为开展高精度水稻种植信息业务化监测提供参考。

Compared with optical remote sensing,synthetic aperture radar(SAR)remote sensing can not be affected by cloud and rain,which provides a new means for accurate monitoring of large-scale crop planting information.Based on the multi-temporal Sentinel-1A SAR image data from 2018 to 2021,a new method for monitoring the planting distribution and area was proposed with Xiaozhan rice in Tianjin as an example,which combined the similarity analysis of growth characteristics with random forest classification.Firstly,the backscattering coefficient time series characteristic curves of different ground objects under VV and VH polarization modes were extracted,and HANTS filtering was used to eliminate the effect of noise.Then,according to the field survey data,the reference growth curve of Xiaozhan rice was obtained and the similarity index of Xiaozhan rice was constructed to screen out the possible planting areas of Xiaozhan rice.Finally,random forest classification model was used to extract the planting area of Xiaozhan rice.The results showed that the multi-temporal Sentinel-1A SAR image similarity analysis combined with random forest classification could obtain high precision rice planting information.The average relative errors of rice planting area extracted by VV and VH polarization methods with the statistical data were 2.67% and 3.80%,respectively.The overall classification accuracies were 95.52% and 93.40%,respectively,and the Kappa coefficients were 0.94 and 0.93,respectively.Compared with the classification results without similarity index,the overall classification accuracy with similarity index under VV and VH polarization modes increased by 4.35 percentage points and 3.13 percentage points,the Kappa coefficients increased by 0.04 and 0.03,and the mapping accuracy of rice increased by 3.38 percentage points and 3.25 percentage points,respectively.The results of this study provide a reference for future business monitoring of high-precision rice planting information.