基于特征优选的多时相SAR数据水稻信息提取方法

多时相合成孔径雷达(Synthetic aperture radar, SAR)数据可为水稻提取提供丰富信息,在多云多雨地区对水稻识别和监测具有独特优势。但过多特征变量的加入,一定程度上造成“维数灾难”及信息冗余,因此,本文提出一种基于多时相后向散射特性及干涉相干性优选特征的水稻提取方法。基于研究区水稻生长周期的多时相Sentinel-1 SAR数据,构建后向散射系数和干涉相干系数特征集,利用ReliefF算法对特征重要性进行排序,同时采用JM距离确定最优特征数目完成最优特征选择,结合随机森林分类算法对研究区水稻进行提取及精度评价。结果表明：基于优选特征提取水稻面积相对误差为4.96%,总体精度达到92.48%,Kappa系数为0.90;从优选特征剔除干涉相干特征提取的水稻面积相对误差增加2.39个百分点,总体分类精度和Kappa系数分别降低4.03个百分点、0.06,说明干涉相干性有利于水稻信息提取。基于多时相后向散射特性及干涉相干性的特征优选减少了数据冗余,提高了运算效率,可实现大范围高精度水稻提取。

Extraction of Rice Information Using Multi-temporal SAR Data Based on Feature Optimization

Synthetic aperture radar (SAR) data has unique advantages for rice identification and monitoring in cloudy and rainy weather. Multi-temporal SAR and multi-features can provide rich information for rice extraction, but too many feature variables will cause dimension disaster and information redundancy to some extent. Therefore, a rice extraction method based on multi-temporal backscattering characteristics and coherent coefficient optimization features was proposed. Based on the multi-temporal Sentinel-1 SAR data during the rice growth cycle in the study area, the feature sets of backscattering coefficient and coherence coefficient were constructed, and the importance of the features was sorted by ReliefF algorithm. At the same time, JM distance was used to determine the optimal number of features to complete the optimal features selection. According to the optimal features, the rice planting area in the study area was extracted by the random forest classification algorithm. The results showed that the error of rice area extraction based on the optimal features was 4.96%, the overall accuracy planting was 92.48%, and the Kappa coefficient was 0.90. Excluding coherence coefficient features from the optimal features to extract rice, the area error was increased by 2.39 percentage points, and the overall classification accuracy and Kappa coefficient were decreased by 4.03 percentage points and 0.06, respectively, which showed that coherence coefficient was beneficial to rice information extraction. Based on the characteristics of multi-temporal backscattering and coherence coefficient, data redundancy was reduced, operation efficiency was improved, and large-scale and high-precision rice extraction can be realized.