Correlation Analysis of Vegetation Drought and Influencing Factors and a Preliminary Prediction Model Based on Convolutional LSTM

Abstract

The frequent occurrence of droughts has seriously affected agricultural production and human life, and prediction of droughts and timely preparation and countermeasures are important to mitigate this natural disaster. In this study, we collected data of elements closely related to vegetation drought: standardized precipitation evapotranspiration index (SPEI), temperature condition index (TCI), soil moisture data (SM), and vegetation condition index (VCI), and unified the data to the same spatial scale, and then used Pearson and Spearman correlation coefficients to calculate the correlations between elements separately and analyzed the correlation degree and time lag effect of them. In addition, SPEI, SM, and TCI data were selected as inputs and VCI as output, and the corresponding time series data sets were produced. The convolutional LSTM model was applied to predict VCI for a certain period of time in the future within mainland China and compared with the actual situation for analysis to verify the effectiveness of the model for vegetation drought prediction. The results show that: 1) the correlation between TCI and VCI data is the highest, the correlations between SM and TCI and VCI are the second highest, and the correlations between SPEI and each other index are the lowest; temperature and soil moisture can effectively influence the development process of vegetation drought. 2)The lag effect between SPEI and SM is obvious, and the lag effect between SM and VCI is not obvious. Meteorological conditions can effectively influence soil moisture within a certain time period, and soil moisture lags behind the effects of precipitation and temperature in time explaining the lag correlation between SPEI and SM. The lagged effects were spatially aggregated, and the lags in the periphery of areas with higher lags were also relatively higher, and the lags in the periphery of areas with lower lags were relatively lower, which verified the spatial propagation law of the drought occurrence mechanism. 3)With the data set time series becoming longer, the effect of the model has a decreasing trend. The expansion of the data set does not necessarily improve the prediction accuracy of the model, and the performance of the model varies for different time periods of data, and the model has time variability. 4) The correlation between the model prediction results and the actual situation is obvious, and the prediction of vegetation drought within the Northeast region of China is better. However, there is variability in the physical geographic elements of each region, and the prediction of the same model for different regions will show geographical variability.