Drought Characteristics And Prediction Models In Northeast China

Drought is one of the most serious natural disasters in the world.The frequent occurrence of droughts has had an extreme adverse impact on the sustainable development of agriculture in our country.The impact of droughts on Northeast area that is the main grain-producing areas of China is particularly serious.Study on drought characteristics and prediction models is one of the important methods for disaster reduction and prevention.Based on the theory of run,this research utilized the China Z Index(CZI),Standardized Precipitation Index(SPI),Standardized Precipitation Evapotranspiration Index(SPEI)and Reconnaissance drought index(RDI)to identify the drought frequency,duration,severity and drought types in Northeast China respectively,and then comparatively analyzed their applicability.Afterwards this research investigated the variation tendency of drought frequency,duration and severity in Northeast China by the Mann-Kendall trend analysis method,the drought return period by Copula function,the periodic change rule of drought characteristics by wavelet analysis;sensitivity of drought to meteorological factors by the extended Fourier Amplitude Sensitivity Test method,respectively;In addition,a variety of drought prediction models based on penalty linear regression and ensemble methods were established and compared,and the optimal drought prediction model based on machine learning methods was selected.Finally,a new Long Short-Term Memory(LSTM)drought prediction model based on deep learning technology was established,and its performance advantages compared with the traditional multi-layer perceptron(MLP)model with similar structure were compared and analyzed,as well as the methods to improve its prediction performance.The main research results are as follows:(1)Study on the applicability of drought index in Northeast China.The frequency,duration and severity of drought based on SPI,SPEI and RDI in Northeast China were relatively consistent in the spatial and temporal distribution characteristics,while SPEI and the corresponding NDVI showed the most significant positive correlation.Furthermore,the SPEI-based drought types were also accord with the drought records over the years in Northeast China that were more spring and autumn droughts,hence SPEI is the most suitable index for evaluating drought conditions in Northeast China.In addition,the drought frequency,duration and severity identified by SPI,SPEI and RDI were influenced significantly by timescales,but their variation characteristics at different timescales were almost agreement.(2)Analysis of the characteristics of the spatial and temporal evolution of drought in Northeast China.The drought frequency and duration in Northeast China showed a downward trend,while the drought severity showed an upward trend,but they were not significant basically.The drought risk was more higher in the western of Northeast China,Heilongjiang Province and northwestern Liaoning,while was lower risk in eastern Liaoning and southern Jilin.The drought frequency,duration and severity showed the periodic characteristics of multi-period superimposed oscillations.The main cycle basically showed a "long-shortlong" change pattern during the study period from 1963 to 2015,and the most significant main cycle was 33-40 a.Drought is most affected by precipitation,but temperature,relative humidity,and wind speed are also important.(3)Study on drought prediction models based on machine learning methods.A variety of drought prediction models based on penalty linear regression and ensemble methods had been established to predict SPEI on 3,6,12 and 24 month timescales.The penalized linear regression models showed better prediction performance than the traditional least square regression(OLS)model,and the Lasso Regression(LR)model had the best prediction result.The ensemble method could improved the prediction accuracy of drought model based on the decision tree(DT),especially the model adopted random forest(RF)had a significant improvement effect.The LR model based on penalty linear regression was fully superior to the RF model in the process of drought prediction,it is therefore the optimal drought model based on machine learning in this study.(4)Study on drought prediction based on deep learning methods.The LSTM model had all good results of predicting SPEI at various timescales,especially suitable for the SPEI at a longer timescale.Compared with the MLP model that had similar structure,it displayed overwhelming superiority and better robustness and generalization ability.The prediction accuracy was slightly improved by adjusting the training parameters and network structure of the LSTM model,but a more complex network structure could reduced the calculation speed instead of improving the prediction accuracy.