集成多源遥感数据与生育期时序光谱特征的水稻种植面积提取

Integration of Multi‑Source Remote Sensing Data and Temporal Spectral Features of Growth Stages for Rice Planting Area Extraction

为快速、有效地提取水稻空间分布信息,利用水稻生育期内的光学和雷达影像,依据水稻不同生育时期内的光谱变化规律,提出了简单实用的时序像元频率约束模型(TPFCM)。首先,采用中值合成法分别将Sentinel-2、Landsat-8和Sentinel-1合成为月度数据,将合成后的Sentinel-2和Landsat-8进行融合以减少云阴影对水稻像元的影响,并选取3种融合影像的特征光谱波段计算归一化植被指数(NDVI)、增强植被指数(EVI)和地表水分指数(LSWI),生成月度多维特征成果影像。其次,检索水稻5个生育时期的遥感影像数据,利用随机森林(RF)分类器提取各生育时期的种植面积,并输入TPFCM模型计算水稻生育期内每个水稻像元的频率,最终依据预提取精度、提取面积误差阈值条件控制模型输出最优水稻空间分布信息。结果表明,TPFCM模型输出的水稻提取面积相对误差为-3.83%,与基于RF分类器的单一生育时期的水稻提取面积相对误差相比绝对值减少了3.94百分点,且提取面积与统计参考面积相关性显著(R^(2)=0.97)。

To fast and effective extraction of rice spatial distribution information,this study proposed a simple and practical time‑series pixels frequency constraint model(TPFCM)based on the spectral variation pattern of rice development using optical and radar images during the growth period of rice.Firstly,the median synthesis method was used to synthesize Sentinel‑2,Landsat‑8,and Sentinel‑1 as monthly data,and the synthesized Sentinel‑2 and Landsat‑8 were fused to reduce the influence of cloud shadow on rice image pixels,and the feature spectral bands of the three fusion images were selected to calculate the normalized vegetation index(NDVI),enhanced vegetation index(EVI)and land surface water index(LSWI)to create monthly multi‑dimensional feature images.Secondly,the random forest classifier was used to initially extract the rice area within the five important development stages,and input into the TPFCM model to calculate the frequency of rice pixels within each stage in the growth period.Finally,the model was controlled to output the optimal rice spatial distribution information based on the pre‑extraction accuracy and area error threshold conditions.The results showed that the TPFCM model output-3.83%of rice planting area error,which was 3.94 percentage points less than that of single‑phase rice extraction area error by comparing the absolute values,and the correlation between the extracted area and statistical reference area was significant(R^(2)=0.97).