基于Sentinel-1数据时序特征的热带地区水稻种植结构提取方法

由于热带地区的雨季时间较长,云覆盖严重,基于光学影像难以准确提取区域内的水稻种植模式。该文以泰国湄南河流域中部平原水稻种植区为例,基于Sentinel-1SAR时间序列数据,提出一种融合时序统计参数与时序曲线相似性特征的热带地区水稻种植结构提取方法。首先利用年内所有可获取的Sentinel-1SAR数据,分别基于像元和基于对象构建后向散射系数时间序列曲线,提取时序特征参数;利用动态时间规整(Dynamic Time Warping,DTW)算法,计算后向散射系数时序曲线与地物标准曲线间的隶属度;将时序特征参数、时序曲线隶属度相结合,利用随机森林模型进行机器学习监督分类,提取研究区的水稻种植信息并评价分类精度。结果表明,基于Sentinel-1SAR时序特征融合的算法可以较好地提高水稻种植结构分类精度。其中,基于对象的分类算法的单季稻提取用户精度为81.46%,生产者精度为82.00%;双季稻用户精度为88.0%,生产者精度为84.08%,均优于基于像元的分类算法。研究结果可为多云多雨的热带地区水稻种植信息提取提供一种新的思路。

Extraction of rice planting structure in tropical region based on Sentinel-1 temporal features integration

Rapid and accurate extraction of rice planting information is of great significance for regional rice planting monitoring,yield evaluation and production management.Thailand is located in the central part of Indo-china Peninsula,with a humid tropical monsoon climate,with an annual average temperature of 27℃and an annual average precipitation of more than 1000 mm.Many areas are suitable for double rice cultivation.However,because of the long rainy season and large amount of cloud,it is difficult to obtain high-quality optical remote sensing images for crop classification.In addition,the diversity of rice planting structure also hinders the accurate recognition of complex rice planting modes based on traditional optical images.In this paper,a multi-feature classification method for rice planting information extraction based on time series Sentinel-1 SAR data was proposed.First,all sentinel-1 SAR data available in a whole year were used to construct the time series profiles of backscatter coefficient at the pixel level and object level,respectively.The backscatter coefficient profiles were de-noised based on Savitzky-Golay filtering algorithm using the TIMESAT software,then the Dynamic Time Warping(DTW)distance-based algorithm at the pixel level(Pixel-Based DTW,PBDTW)and object level(Object-Based DTW,OBDTW)were applied to measuring the similarity of backscatter coefficient profiles between the target land classes and reference land classed.Furthermore,the max value,min value,mean and standard deviation of the backscatter coefficient were calculated.The time series statistical feature parameters were then integrated with membership features for Random Forest classification,and the performance of different combinations were assessed based on classification confusion matrix.The results showed that backscatter coefficient profile was an effective way to represent the phenological information contained in time-series Sentinel-1 SAR data.By matching the similarity of time series profiles,single rice and double rice could be well identified from other crops.After adding the time series statistical feature parameters,the user’s accuracy and the producer’s accuracy of PBDTW algorithm increased by 6.62 and 6.76 percentage points for single rice,and by 5.34 and 3.66 percentage points for the double rice.Compared with the OBDTW algorithm only,the user’s accuracy and the producer’s accuracy of OBDTW combined with time series statistical feature parameters algorithm increased by 5.3 and 4.82 percentage poins for single rice,and 3.34 and 5.46 percentage points for double rice.The results also indicated that OBDTW algorithm could reduce the influence of noise by calculating the average value of backscatter coefficients of all pixels belonging to the object,so the classification accuracy of OBDTW algroithm was higher than that of PBDTW algorithm.The combination of OBDTW together with time series statistical feature parameters had the highest classification accuracy,with the user’s accuracy 81.46%and producer’s accuracy 82.00%for single rice,and 86.87%and 84.08%for double rice,respectively.The results can provide a new way to extract rice planting information in the cloudy and rainy tropics.