Leaf area index (LAI), indicating the area of leaves per unit of ground surface area, characterizes the surface area with which energy, mass and momentum are exchanged between the vegetated land surface and the planetary boundary layer. Changes in LAI impact climate, the terrestrial carbon cycle, and the atmospheric chemistry through the emission and deposition of several compounds. LAI in cropland is also a good indicator of crop yield. Considering LAI’s important role in indicating vegetation growth, short-term prediction of LAI is of great importance for both cropland and natural ecosystems. Previous work in predicting LAI mainly used statistical equations or traditional machine learning approaches like random forest and artificial neural network, few deep learning models have been applied in this area. Among all deep learning methods, recurrent neural network (RNN) and long short term memory (LSTM) have been proved successful in research tasks such as time series prediction. And LSTM is considered more effective than RNN because it contains memory cell in the recurrent hidden layer to keep track of long time and short time input. Thus, we used LSTM in this study to make prediction of LAI in different latitudes and land cover types.
Results/Conclusions
We firstly calculated the mean LAI value for each land cover types, then, used LSTM to train and predict LAI. We set 300 months as training period (1982-2006) and 120 months as validating period (2007-2016). According to the results, 6 out of 11 land cover types have very high prediction accuracy (R2>0.94). LSTM’s good performance in predicting LAI in cropland implies its potential in crop yield prediction. However, LSTM’s prediction accuracy is low in deciduous broadleaf forest, evergreen broadleaf forest, and woody savanna (R2<0.62). In general, the accuracy of LSTM in predicting LAI is high (R2>0.8) across most latitudes and land cover types. The relative low prediction accuracy in deciduous broadleaf forest, evergreen broadleaf forest, and woody savanna and tropical regions may be induced by the low quality of satellite-based LAI products and insufficient consideration of explanatory variables of LAI. Experiments with different amounts of training data suggest that the increase of training data can effectively improve the prediction accuracy of LSTM, especially for regions where LSTM has relative poor performance.