Spatial Statistical Modeling And Prediction Via Gradient Boosting Learning

Climate change is seriously affecting the environment,hydrology and ecosystems,obtaining accurate high-resolution data is important for studying climate change,especially in countries like Bangladesh,where meteorological stations are scarce.Statistical downscaling is an effective method for obtaining high spatial resolution climate data,which is characterized by low computational complexity,easy construction of models and numerous methods,compared with the traditional statistical downscaling method,which hardly reveal the complicated nonlinear relationship between local variables and large-scale variables,machine learning algorithm performs well in solving complex nonlinear correlation problems among variables.Based on daily precipitation and temperature data from the meteorological stations in Bangladesh from 1989 to 2018,this study developed a downscaling model based on three machine learning algorithms(Support Vector Machine,Random Forest and Gradient Boosting Regressor),compared with the traditional multiple linear regression downscaling model,the results show that the downscaling model based on machine learning is obviously superior to the traditional multiple linear regression downscaling model.Compared with Random Forest(R2=0.94),Support Vector Machine(R2=0.88)and Multiple Linear Regression(R2=0.69)models,the Gradient Boosting Regressor had the best performance,with R2 as high as 0.98,the effect is improved by 42%than traditional multiple linear regression downscaling model.Based on the optimal downscaling model,the spatial distribution maps of high spatial resolution temperature and precipitation in Bangladesh from 1989 to 2018 are obtained,the drought is expected to continue to worsen owing to overall high temperatures and low precipitation in the western and northern parts of Bangladesh.The study shows that there is a statistically significant spatial variability in the distribution of precipitation and temperature in Bangladesh,the precipitation in the eastern,southwestern,southern and southeastern regions of Bangladesh show a statistical upward trend,passing the 90%significance test,while the precipitation in the northern,northwestern and central regions show a statistical downward trend.The mean annual precipitation in Bangladesh ranges from22.9 to 26.0 mm/year,with the northern and central regions showing the largest decreasing trend and most regions showing no significant decreasing trend.The temperature in the north,east and south regions show a statistically significant increasing trend,which passed the significance test of 99%.The mean annual temperature increases in the northern region is as high as 0.0165℃,the mean annual temperature warming trend ranges from 0.0052 to 0.03℃/year.Based on high spatial resolution temperature and precipitation data from 1989 to 2018,the XGBoost prediction model is established to obtain the temperature and precipitation data from 2025 to 2035 in Bangladesh.The performance of the XGBoost model in temperature prediction was better than that precipitation,with R2 reaching 0.94.After empirical orthogonal decomposition,the distribution of the first group of annual precipitation modes presents the characteristics of the southwest-northeast inverse phase pattern,the second group of modes shows a north-south "+,-,+" tripolar distribution,and the third group of modes is consistent with the first group of modes,indicating that future precipitation distribution in Bangladesh as a whole will be more balanced.The spatial distribution of the first group of temperature modes is characterized by the south-north inverse phase pattern and the second group by empirical orthogonal decomposition,the spatial distribution of the third group of modes shows the characteristics of the east-west inverse phase distribution.The southeast of the three modes is the phase center region,which indicates that the region is sensitive to temperature change.