A gap filling method for daily evapotranspiration of global flux data sets based on deep learning

Abstract

In response to irregular data gaps in evapotranspiration (ET) data obtained using eddy covariance (EC) methods, this study seeks to explore a high-precision interpolation method that is not constrained by spatiotemporal limitations. Leveraging daily observations from 241 global flux towers (FLUXNET ET) as a reference, data gaps are artificially introduced at five different proportions: 5 %, 10 %, 20 %, 35 %, and 50 %, this process is repeated 30 times (Take the average of the results). Four ET products, along with environmental factors (ERA5 meteorological data and NOAA Normalized Difference Vegetation Index), are employed as input to assess the performance of the target model, Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM), as well as five comparative models: CNN, LSTM, Random Forest (RF), Support Vector Machine (SVM), and Extreme Learning Machine (ELM), in filling the data gaps. These results are then compared with the filling outcomes obtained without ET products as input, aiming to investigate whether ET products contributes to improving the accuracy and stability of ET gap filling. The results indicate that, firstly, the filling performance of the target model CNN-LSTM surpasses that of other models, with minimal impact from gap proportions. Moreover, incorporating ET products as input alongside environmental factors significantly enhances the filling accuracy and stability of the model. Compared to scenarios without ET products as input, the CNN-LSTM model shows an average increase in R2 of 5 % to 15 % and a decrease in root mean square error (RMSE) of 15 % to 35 %. Specifically, as the gap proportion increases from 5 % to 50 %, the average Pearson correlation coefficient (R2) decreases from 0.952 to 0.846 (a decline of 11 %), the average RMSE increases from 0.491 mm/d to 0.620 mm/d (a rise of 26 %), and reached the filling limit at the gap ratio of 93 % (R2 below 0.4). Conversely, without ET products as input, as the gap proportion increases, the average R2 decreases from 0.916 to 0.784 (a decline of 14 %), the RMSE increases from 0.552 mm/d to 0.952 mm/d (a rise of 72 %), and reached the filling limit at the gap ratio of 80 %. Secondly, the contribution of ET products to the model estimation far exceeds that of ERA5 meteorological factors and NDVI, with the contribution of ET products increasing with higher gap proportions from 58.0 % to 73.6 %. Finally, different latitudes, land cover types, and climate types influence the model filling performance, with better performance observed in high-latitude regions, Grassland (GRA), and Dry-Cold climates. In summary, incorporating ET products as input for CNN-LSTM yields high-precision gap-filling results, surpassing existing typical methods for gap filling and its application is not constrained by spatiotemporal limitations.