An enhanced spatiotemporal fusion method – Implications for DNN based time-series LAI estimation by using Sentinel-2 and MODIS

• DOI: 10.1016/j.fcr.2022.108452
• Publication Date: 2022-02-02T00:04:17Z
• Authors: Yan Li; Yanzhao Ren; Wanlin Gao; Jingdun Jia; Sha Tao; Xinliang Liu
• Keywords: Robustness

The consequent and accurate monitoring of the seasonal dynamics crop leaf area index (LAI) is critical to yield estimation agriculture policy development. It difficult for a single sensor balance spatial temporal resolution. Spatiotemporal fusion an effective way meet need high applications. Among methods, regression model Fitting, Filtering residual Compensation (Fit-FC) may be recommended vegetation dynamic because its outperformance cases with considerable phenological changes. However, it not good at capturing image structure textures. To overcome limitations, enhanced version Fit-FC, referred as Enhanced-Fit-FC (EFF), was developed. EFF method can applied one (unidirectional prediction, Uni-EFF) or two (bidirectional Bi-EFF) coarse-fine pairs input near real-time post-growth visual quantitative assessments indicated that mitigated blurring effect Fit-FC generated reflectance, especially Bi-EFF contributed land cover Compared correlation coefficient (CC) quality (QI) increased by more than 0.12, root mean square error (RMSE) decreased 0.16 maximum. Further, we identified robustness adaptability deep neural network (DNN) time-series LAI estimation. results substantiated effectiveness DNN in dealing nonlinear problems alleviating spectral saturation higher CC lower RMSE relative (rRMSE) whole growth-stages (CC= 0.91, RMSE= 0.28, rRMSE= 7.18%) vegetative stage (CC=0.94, RMSE=0.24, 5.81%). In conclusion, proposed this study competent constructing synthetic images Moreover, shows potential growth-stages. This research only contributes dense remote sensing-based applications (e.g., change monitoring, forecast), but also valuable high-spatial-precision farmland management.