A methodology to reconstruct LAI time series data based on generative adversarial network and improved Savitzky-Golay filter

Abstract

High-quality leaf area index (LAI) data is essential for regional and global ecology, climate and environment research. However, there are still many quality problems in the continuity of current LAI time series products. Here we developed a new comprehensive three-step reconstruction method (GANSG) for satellite-retrieved LAI time series data based on generative adversarial network, improving the Savitzky-Golay (S-G) filter and median absolute deviation filter. We applied GANSG to the reconstruction of MODIS LAI data in China from 2001 to 2019. The reconstruction results show that the new method based on the unsupervised deep learning framework has an advantage in interpolating low-quality LAI with high precision. The new method can better retain high-quality pixel information to smoothen the interpolated LAI time series by improving the traditional S-G filter. Compared with the five other methods, including the adaptive S-G filter, double logistic, asymmetric Gaussian, modified temporal spatial filter and spatial temporal S-G filter, qualitative analysis showed the new method has a more resilient ability to handle the continuous loss of high-quality pixels and identify the phenological features of biomes. Quantitative analysis based on station observation showed that the new method performs best among the other three methods, with the optimal correlation coefficient of 0.84 relative to station observation and the lowest root mean square error of 0.71 m2/m2.