基于多时相遥感影像的水稻种植信息提取

Rice Planting Information Extraction Based on Multi-temporal Remote Sensing Images

获取水稻种植信息对于指导水稻生产,监测作物生长及合理分配水资源具有重要意义。针对基于单时相影像提取水稻信息精度有限,以Sentinel-2A/B多时相影像为数据源,构建NDVI、EVI、NDWI和光谱特征4种时序特征数据集并设计6种试验方案,结合随机森林算法对水稻种植信息进行提取。结果表明,NDVI、EVI时序曲线可以较好反映出水稻生育期的物候特征,不同地类的光谱时序曲线和NDWI时序曲线可分离度较高,有利于提高分类精度;基于NDVI时序数据集的分类精度最低,基于光谱时序数据集的分类精度最高,总体精度达95.5590%,Kappa系数为0.9433,与基于NDVI的分类结果相比,总体精度、Kappa系数、水稻生产者精度和用户精度分别提高了3.5304%、0.0449、8.64%和3.36%,水稻与旱地的混分现象得到有效抑制。该研究为区域水稻种植信息精确提取在数据源选择、时序特征构建方面提供了一种新的思路和技术手段。

Obtaining rice planting information is of great significance for guiding rice production,monitoring crop growth and rational allocation of water resources.In view of the limited accuracy of extracting rice information based on single temporal image,four time-series feature data sets of NDVI,EVI,NDWI and spectral features were created based on sentinel-2A/B multi-temporal images.Six experimental schemes were designed to extract rice planting information combined with random forest algorithm.The results showed that NDVI and EVI time series could better reflect the phenological characteristics of rice growth period,and the spectral time series and NDWI time series of different land types had a high degree of separation,which was conducive to improve the classification accuracy;the classification accuracy based on NDVI time series dataset was the lowest,and the classification accuracy based on spectral time series dataset was the highest,the overall accuracy was 95.5590%,and the Kappa coefficient was 0.9433.Compared with the classification results based on NDVI,the overall accuracy,Kappa coefficient,rice producer accuracy and user accuracy were improved by 3.5304%,0.0449,8.64% and 3.36%,respectively.And the mixing of rice and dry land was effectively controlled.This research provided a new idea and technical means for accurate extraction of regional rice planting information in data sources selection and time series feature construction.