1/2数据特征优选的冬小麦种植区识别方法研究

为了提高冬小麦种植区识别精度，本文基于谷歌地球引擎（Google Earth Engine，GEE）平台和随机森林算法，对比雷达和光学遥感数据对冬小麦提取效果的差异，并对多类特征变量进行重要性分析，研究特征优选对冬小麦识别精度的影响。选取2019年3—5月冬小麦关键生育期的Sentinel-1和Sentinel-2影像为数据源，构建Sentinel-1的极化特征和纹理特征以及Sentinel-2的光谱特征、植被指数特征、植被指数变化率特征共5类特征变量；设置不同数据源和不同特征组合的冬小麦种植区提取方案；对方案中特征变量进行优选，得出最优特征组合，利用最优特征组合对河南省驻马店市冬小麦种植区进行提取。结果表明，无论是否进行特征优选，基于多源遥感数据的冬小麦识别精度均优于仅采用光学或雷达数据的精度；经过特征优选后，各方案的分类精度均有不同程度的提升，说明多源数据特征变量组合和特征优选均能够提高分类精度。不同月份和类型的特征变量对分类精度的贡献率不同，贡献率由大到小为4月、3月和5月；贡献率由大到小的特征类型为极化特征、植被指数变化率特征、植被指数特征、光谱特征和纹理特征。基于多源数据特征优选提取的2019年驻马店冬小麦空间分布最优，总体精度为95.60%，Kappa系数为0.93，冬小麦提取面积与统计年鉴数据相比，相对误差为2.23%。本文可为基于多源光学和雷达遥感影像进行农作物种植区提取的研究提供理论参考。

Abstract:

In order to improve the accuracy of winter wheat identification, the difference between radar and optical remote sensing data on winter wheat area extraction was compared and analyzed based on Google Earth Engine (GEE) platform and random forest algorithm. The importance analysis of multiple feature variables was performed to study the influence of feature optimization on the accuracy of winter wheat extraction. The Sentinel-1 and Sentinel-2 images during the main growth period of winter wheat (from March 1 to May 31, 2019) were chosen as the data sources. The polarization and texture features of Sentinel-1 data as well as the spectral, vegetation index and vegetation index change rate features of Sentinel-2 data were constructed. Six winter wheat identification schemes were constructed based on different remote sensing data sources and feature combinations, and the accuracies of the schemes were compared and analyzed. Then the feature variables were optimized and the optimal feature combination was obtained to extract the planting area of winter wheat in Zhumadian City, Henan Province. The results showed that regardless of feature optimization, the results of winter wheat area extraction based on multi-source remote sensing data were superior to those by using only optical or radar data. After feature optimization, the classification accuracy of each scheme was further improved, indicating that both the combination of multi-source feature variables and feature optimization can improve the winter wheat identification accuracy. In addition, the feature variables of different months and types had different contribution rates to classification accuracy, and the months with contribution rates from high to low were April, March and May. The feature types with contribution rates from high to low were polarization, vegetation index change rate, vegetation index, spectral features and texture. The accuracy of winter wheat extraction in Zhumadian based on both multi-source satellite data and feature optimization were the best, with the overall accuracy of 95.60% and Kappa coefficient of 0.93. The relative error between the extracted area of winter wheat and official statistical data was 2.23%. The research result can provide an important theoretical reference for crop planting area extraction based on multi-source optical and radar remote sensing images.