基于Sentinel的时间序列田块尺度LAI重建与冬小麦估产

Reconstruction of Time Series LAI and Winter Wheat Yield Estimation at Field Scales Based on Sentinel Satellites

为了进行田块尺度的冬小麦单产估测,以陕西省关中平原为研究区域,基于Sentinel-1、Sentinel-2和Sentinel-3卫星数据反演叶面积指数(LAI),并利用增强的深度卷积神经网络融合模型(EDCSTFN)和增强的时空自适应反射率融合模型(ESTARFM)对Sentinel-1、Sentinel-2和Sentinel-3 LAI进行时空融合,进而重建尺度12 d的空间分辨率20 m LAI并用于冬小麦单产估测。结果表明,基于Sentinel-1后向散射系数和相干性能够准确地反演关中平原冬小麦种植区的20 m空间分辨率LAI,决定系数(R^(2))在冬小麦主要生育期可达0.70以上;相比于基于Sentinel-2和Sentinel-3的ESTARFM模型和EDCSTFN模型(EDCSTFN_S3),基于Sentinel-1和Sentinel-2的EDCSTFN模型(EDCSTFN_S1)可以明显提高距离参考影像获取日期较远的日期的LAI时空融合精度,ESTARFM、EDCSTFN_S3和EDCSTFN_S13个模型在5月下旬的融合结果对应的R^(2)分别为0.53、0.71和0.76;基于时空融合LAI的冬小麦估产结果与冬小麦单产数据具有良好的相关性(R^(2)=0.52,P<0.01),估产结果的均方根误差为358.25 kg/hm~2,归一化均方根误差为19%,平均相对误差为7.34%,并显示了丰富的田块尺度冬小麦单产分布细节特征,展现了进行田块尺度冬小麦精确估产的潜力。

Time-series crop growth monitoring at field scales is very important for crop management and yield estimation.However,the spatial resolutions(250~1000 m)of current satellite sensors with high temporal resolutions are too coarse for areas with complex and diverse land-use types,such as the Guanzhong Plain of Shaanxi Province,which causes great uncertainties in crop yield estimation results.To estimate the winter wheat yield at field scales,a study was carried out on the Guanzhong Plain.The enhanced deep convolutional spatiotemporal fusion network(EDCSTFN)model and enhanced spatial and temporal adaptive reflectance fusion model(ESTARFM)were used to fuse the leaf area index(LAI)retrieved from Sentinel-1,Sentinel-2 and Sentinel-3 imagery,thereby reconstructing the LAI imagery with a 20 m spatial resolution at 12-day interval.Finally,the reconstructed LAI imagery were used for the winter wheat yield estimation at field scales.The results showed that the LAI imagery with a 20 m spatial resolution in the winter wheat planting area of Guanzhong Plain can be accurately retrieved based on the backscatter coefficient and coherence derived from Sentinel-1 data,and the coefficient of determination(R^(2))in March and April were larger than 0.70.Compared with the ESTARFM and EDCSTFN models based on Sentinel-2 and Sentinel-3(EDCSTFN_S3),the EDCSTFN model based on Sentinel-1 and Sentinel-2(EDCSTFN_S1)can significantly improve the accuracy of LAI spatio-temporal fusion results on the dates far from the reference dates(R^(2)values for ESTARFM,EDCSTFN_S3 and EDCSTFN_S1 model in late May were 0.53,0.71 and 0.76,respectively).The winter wheat yield estimation results based on the spatio-temporal fused LAI had a good correlation with the winter wheat yield data(R^(2)=0.52,P<0.01),the root mean square error of the yield estimation results was 358.25 kg/hm~2,the normalized root mean square error was 19%and the average relative error was 7.34%.In addition,the yield estimation results showed more spatial distribution details of winter wheat yield at field scales,thereby indicating the potential for accurate winter wheat yield estimation at field scales.